Article

Dynamic Monitoring and Precision Fertilization Decision

System for Agricultural Soil Nutrients Using UAV Remote
Sensing and GIS
Xiaolong Chen, Hongfeng Zhang and Cora Un In Wong *

Faculty of Humanities and Social Sciences, Macao Polytechnic University, Macao, China;
[email protected] (H.Z.)
* Correspondence: [email protected]

Abstract
We propose a dynamic monitoring and precision fertilization decision system for agricul-
tural soil nutrients, integrating UAV remote sensing and GIS technologies to address the
limitations of traditional soil nutrient assessment methods. The proposed method com-
bines multi-source data fusion, including hyperspectral and multispectral UAV imagery
with ground sensor data, to achieve high-resolution spatial and spectral analysis of soil
nutrients. Real-time data processing algorithms enable rapid updates of soil nutrient status,
while a time-series dynamic model captures seasonal variations and crop growth stage
influences, improving prediction accuracy (RMSE reductions of 43–70% for nitrogen, phos-
phorus, and potassium compared to conventional laboratory-based methods and satellite
NDVI approaches). The experimental validation compared the proposed system against
two conventional approaches: (1) laboratory soil testing with standardized fertilization
recommendations and (2) satellite NDVI-based fertilization. Field trials across three distinct
agroecological zones demonstrated that the proposed system reduced fertilizer inputs by
18–27% while increasing crop yields by 4–11%, outperforming both conventional methods.
Furthermore, an intelligent fertilization decision model generates tailored fertilization plans
by analyzing real-time soil conditions, crop demands, and climate factors, with continu-
Academic Editor: Yongchao Tian
ous learning enhancing its precision over time. The system also incorporates GIS-based
visualization tools, providing intuitive spatial representations of nutrient distributions
Received: 15 June 2025
Revised: 18 July 2025
and interactive functionalities for detailed insights. Our approach significantly advances
Accepted: 23 July 2025 precision agriculture by automating the entire workflow from data collection to decision-
Published: 27 July 2025 making, reducing resource waste and optimizing crop yields. The integration of UAV
Citation: Chen, X.; Zhang, H.; Wong, remote sensing, dynamic modeling, and machine learning distinguishes this work from
C.U.I. Dynamic Monitoring and conventional static systems, offering a scalable and adaptive framework for sustainable
Precision Fertilization Decision farming practices.
System for Agricultural Soil Nutrients
Using UAV Remote Sensing and GIS.
Keywords: UAV remote sensing; GIS technology; soil nutrient monitoring; precision
Agriculture 2025, 15, 1627.
fertilization; multi-source data fusion; dynamic modeling
https://doi.org/10.3390/
agriculture15151627

Copyright: © 2025 by the authors.

Licensee MDPI, Basel, Switzerland.
1. Introduction
This article is an open access article
distributed under the terms and Modern agriculture faces increasing demands for sustainable and efficient production
conditions of the Creative Commons systems, with soil nutrient management being a critical factor influencing crop yield and
Attribution (CC BY) license environmental sustainability. Traditional soil nutrient monitoring methods rely heavily
(https://creativecommons.org/
on field sampling and laboratory analysis, which are time-consuming and costly, and
licenses/by/4.0/).

Agriculture 2025, 15, 1627 https://doi.org/10.3390/agriculture15151627

Agriculture 2025, 15, 1627 2 of 27

lack spatial continuity [1–3]. Remote sensing technologies have emerged as promising
alternatives, offering non-destructive and large-scale monitoring capabilities. Satellite-
based multispectral and hyperspectral imaging have been widely used for soil property
estimation [4–6], yet their spatial and temporal resolutions often prove inadequate for
precision agriculture applications.
The advent of Unmanned Aerial Vehicle (UAV) remote sensing has revolutionized
agricultural monitoring by providing high-resolution imagery with flexible acquisition
schedules [7]. UAV platforms equipped with multispectral and hyperspectral sensors
have demonstrated remarkable potential in crop health assessment and soil property es-
timation [8,9]. When combined with Geographic Information System (GIS) technologies,
these data enable sophisticated spatial analysis and visualization of soil nutrient distribu-
tions [10]. However, existing systems often operate in isolation, failing to fully exploit the
synergies between different data sources and analytical approaches.
Recent advances in data fusion techniques effectively integrate multi-source agricul-
tural data, including satellite imagery, UAV data, and ground sensor measurements [11,12].
While machine learning algorithms successfully process these datasets for crop yield pre-
diction [13], most current systems use static models. These models fail to capture dynamic
nutrient variations caused by seasonal changes, crop growth stages, and environmental
factors [14,15].
Existing soil nutrient monitoring systems often operate in isolation, failing to fully
leverage the complementary strengths of different analytical approaches. Conventional
methods relying solely on satellite remote sensing face inherent resolution limitations that
compromise their precision for field-scale applications, while laboratory-based soil testing,
despite its accuracy, cannot provide the spatial continuity or temporal responsiveness
required for dynamic agricultural management. Furthermore, traditional static models
struggle to adapt to the rapidly changing nutrient conditions influenced by environmental
factors and crop growth stages. These fundamental limitations in current approaches
highlight the critical need for an integrated system that combines multi-source data fusion
with real-time processing capabilities and adaptive modeling techniques to enable truly
precision agriculture.
We propose a novel integrated system that addresses these limitations through three
key innovations. First, our method combines UAV remote sensing with ground sensor
data in a multi-source fusion framework, achieving unprecedented spatial and spectral
resolution in soil nutrient monitoring. Second, we develop a dynamic modeling approach
that continuously updates soil nutrient status based on time-series analysis, accounting for
temporal variations in nutrient availability. Third, we implement an intelligent decision-
making module that generates precision fertilization recommendations by analyzing real-
time soil conditions, crop requirements, and environmental factors, with machine learning
algorithms continuously improving prediction accuracy.
The proposed system differs from conventional approaches in several important as-
pects. Unlike traditional laboratory-based methods, our UAV-GIS integration enables rapid,
large-scale monitoring with high spatial resolution. Compared to existing remote sensing
systems, our dynamic modeling approach captures temporal variations more effectively.
Furthermore, our intelligent decision module goes beyond static recommendation systems
by incorporating continuous learning and real-time data processing capabilities [16].
This paper makes four main contributions to the field of precision agriculture. We
present (1) a comprehensive UAV-GIS framework for high-resolution soil nutrient moni-
toring, (2) a dynamic modeling approach for time-series nutrient status prediction, (3) an
intelligent fertilization decision system with continuous learning capabilities, and (4) ex-
perimental validation demonstrating the system’s effectiveness in real-world agricultural
Agriculture 2025, 15, 1627 3 of 27

scenarios. Our approach represents a significant advancement over existing methods

by integrating the entire workflow from data collection to decision-making in a unified,
automated system.
The remainder of this paper is organized as follows: Section 2 reviews related work in
soil nutrient monitoring and precision fertilization systems. Section 3 provides background
knowledge on UAV remote sensing and GIS technologies. Section 4 details our proposed
methodology, while Section 5 describes the experimental setup. Results and analysis are
presented in Section 6, followed by discussion and future work in Section 7. The paper
concludes with Section 8.

2. Related Work
Recent advancements in precision agriculture have seen significant developments
in soil nutrient monitoring and fertilization decision-making systems. These approaches
can be broadly categorized into three main research directions: remote sensing-based soil
nutrient estimation, data fusion techniques for agricultural monitoring, and intelligent
fertilization decision systems.

2.1. Remote Sensing for Soil Nutrient Monitoring

Satellite remote sensing has long been employed for large-scale soil property esti-
mation, with MODIS and Landsat data being particularly popular due to their extensive
temporal coverage [17,18]. However, these platforms suffer from inherent limitations
in spatial resolution, typically ranging from 10 to 250 m, which proves insufficient for
field-scale precision agriculture applications. More recent studies have explored the po-
tential of Sentinel-2 data, which offers improved spatial resolution (10–60 m) and spectral
characteristics suitable for soil organic matter estimation [19].
Recent work by Yuan et al. [20] and Jiang et al. [21] has enhanced UAV-based soil
nutrient prediction using deep learning, achieving higher spatial resolution than satellite-
based methods. Their studies demonstrate that UAV-mounted hyperspectral sensors can
now detect nitrogen variations at 2 cm resolution when combined with convolutional
neural networks, representing a significant leap forward in precision soil monitoring.
The emergence of UAV-based remote sensing has addressed many of these reso-
lution limitations. Researchers have demonstrated the effectiveness of UAV-mounted
multispectral sensors in estimating soil nitrogen content with sub-meter resolution [22,23].
Hyperspectral UAV systems have shown even greater promise, with spectral resolutions
sufficient to identify subtle variations in soil phosphorus and potassium levels [24]. These
airborne systems provide the flexibility of on-demand data acquisition, enabling timely
monitoring aligned with critical crop growth stages.

2.2. Multi-Source Data Fusion in Agriculture

The integration of remote sensing data with ground-based measurements has emerged
as a powerful approach to improve soil nutrient estimation accuracy. Several studies
combine satellite imagery with proximal soil sensing data, improving soil property predic-
tion [25,26]. Advanced systems integrate weather station data and historical crop yields to
create detailed soil fertility maps [27].
Recent work has particularly focused on fusing UAV remote sensing with IoT sensor
networks deployed in agricultural fields [28,29]. This combination provides both high-
resolution spatial data from UAVs and continuous temporal monitoring from ground
sensors. Machine learning techniques, especially ensemble methods and deep learning,
have proven effective in processing these heterogeneous datasets to generate accurate soil
nutrient predictions [30,31].
Agriculture 2025, 15, 1627 4 of 27

2.3. Intelligent Fertilization Decision Systems

Traditional fertilization recommendation systems have largely relied on static models
based on soil test results and generalized crop nutrient requirements [32]. These approaches
fail to account for dynamic factors such as weather conditions, crop growth stage, and
spatial variability within fields.
More sophisticated systems have incorporated real-time sensor data and weather
forecasts to adjust fertilization recommendations. Some notable implementations include
model-based systems that simulate crop nutrient uptake under varying environmental
conditions [33]. Others have employed rule-based expert systems that codify agronomic
knowledge into decision-making algorithms [34,35].
Recent advancements have seen the application of machine learning to fertilization
decision-making. These systems learn from historical data to predict optimal nutrient
application rates, with some implementations showing promising results in rice and wheat
production systems [36]. Reinforcement learning approaches have been particularly suc-
cessful in adapting recommendations based on continuous feedback from crop response
and environmental changes [37].
The proposed system advances beyond these existing approaches by integrating
three key innovations: (1) a dynamic monitoring framework combining UAV remote
sensing with ground sensor networks, (2) a time-series modeling approach that captures
soil nutrient variations across seasons and growth stages, and (3) an intelligent decision
system that continuously learns from new data to improve fertilization recommendations.
Unlike previous systems that operate with static models or limited data sources, our
approach provides a comprehensive solution that adapts to changing field conditions
while maintaining high spatial and temporal resolution. This integration of cutting-edge
technologies addresses critical gaps in current precision agriculture systems, particularly in
handling the dynamic nature of soil–plant–atmosphere interactions.
This comparative analysis highlights how our system advances beyond existing ap-
proaches through its integrated dynamic monitoring, time-series modeling, and continuous
learning capabilities.

3. Background and Preliminary Knowledge

Understanding the fundamental concepts of soil nutrient dynamics, remote sensing
principles, and GIS technologies is essential for developing an effective precision fertil-
ization system. These interconnected domains form the theoretical foundation for our
proposed methodology.

3.1. Soil Nutrient Dynamics

The availability and transformation of soil nutrients follow complex biogeochemical
cycles that directly influence crop growth and fertilizer requirements. Nitrogen, phospho-
rus, and potassium—the primary macronutrients—exhibit distinct behaviors in agricultural
soils. The nitrogen cycle involves multiple transformations, including fixation, mineraliza-
tion, nitrification, and denitrification processes (Equation (1)).

Fixation Nitri f ication

N2 → NH4+ → NO3− (1)

Soil moisture, temperature, microbial activity, and crop uptake collectively determine
nutrient availability [38]. Cation exchange capacity (CEC) plays a crucial role in retaining
positively charged nutrients like potassium (K+ ) and ammonium (NH4+ ), while phosphorus
availability depends largely on soil pH and mineral composition [39,40]. These dynamic
Agriculture 2025, 15, 1627 5 of 27

interactions create spatial and temporal variability that challenges traditional uniform
fertilization approaches.

3.2. Principles of Remote Sensing

Optical remote sensing measures the interaction between electromagnetic radiation
and surface materials. The reflectance spectrum of soil contains characteristic absorption
features (Equation (2)) related to organic matter content, moisture, and mineral composition:
Re f lected Radiation
Re f lectance = (2)
Incident Radiation
Vegetation indices like NDVI (Normalized Difference Vegetation Index) have been
widely used for crop health assessment [41–43], while newer indices specifically target soil
properties. Hyperspectral sensors capture continuous spectral signatures with hundreds
of narrow bands, enabling detection of subtle nutrient-related spectral features [44,45].
The spatial resolution of UAV-based systems typically ranges from 1 cm to 1 m, allowing
detailed mapping of within-field variability [46].

3.3. Geographic Information Systems (GIS)

GIS provides powerful tools for spatial data integration (Equation (3)), analysis, and
visualization of agricultural information. The core functionality involves:

Spatial Analysis = f (Spatial Data, Analysis Techniques) (3)

Geostatistical methods like kriging interpolate point measurements into continu-

ous surfaces, while zonal statistics summarize data within management zones [47,48].
Topographic features derived from digital elevation models (DEMs) influence nutrient
distribution through water movement and erosion processes [48]. The integration of remote
sensing data with GIS enables generation of prescription maps that guide variable-rate
fertilizer application [49,50].

4. Methodology for Dynamic Monitoring and Precision Fertilization

Our methodology combines UAV remote sensing, GIS technologies, and machine
learning in a comprehensive framework. This system addresses traditional soil nutrient
assessment challenges through four interconnected components.

4.1. Multi-Source Data Fusion and Processing

The data acquisition module combines hyperspectral and multispectral UAV imagery
with ground sensor measurements to achieve comprehensive soil nutrient characterization.
Hyperspectral data provides detailed spectral information across hundreds of narrow
bands, enabling detection of subtle nutrient-related absorption features. The spectral
response R(λ) at wavelength λ is modeled as (Equation (4)):
n
R(λ) = ∑ w i Si ( λ ) + ϵ ( λ ) (4)
i =1

where Si (λ) represents the spectral signature of soil component i, wi denotes its relative
contribution, and ϵ(λ) accounts for measurement noise. Recent advances by Peng et al. [51]
and Lambertini et al. [52] have demonstrated that adaptive feature selection in multi-
sensor fusion can improve nutrient prediction accuracy by 15–22% compared to static
weighting methods. Their hybrid CNN-LSTM architectures specifically address spectral–
spatial–temporal correlations in agricultural data. Multispectral data offers higher spatial
resolution with broader spectral bands, complementing the hyperspectral information.
Agriculture 2025, 15, 1627 6 of 27

Ground sensors provide continuous measurements of soil moisture, temperature, and

electrical conductivity, serving as validation references for the remote sensing data.
The fusion process employs an adaptive weighting scheme (Equation (5)) that dynam-
ically adjusts the contribution of each data source based on measurement conditions and
sensor characteristics. The fused dataset F ( x, y) at location ( x, y) is computed as:

F ( x, y) = αH ( x, y) + βM ( x, y) + γG ( x, y) (5)

where H, M, and G represent hyperspectral, multispectral, and ground sensor data respec-
tively, with α, β, and γ being their corresponding adaptive weights. These weights are
optimized through a machine learning approach that minimizes the reconstruction error
against ground truth measurements.
The dynamic weight allocation mechanism optimizes the data fusion coefficients
in Equation (5) through constrained minimization (Equation (6)). As detailed in Algo-
rithm 1, the optimization procedure iteratively adjusts the weights of hyperspectral (H),
multispectral (M), and ground sensor (G) data sources to minimize reconstruction error.

arg min ∥ Ntrue − (αH + βM + γG ) ∥2

α,β,γ (6)
subjectto : α + β + γ = 1, 0 ≤ α, β, γ ≤ 1

The adaptive weighting process in Equations (5) and (6) operates through an iterative
optimization procedure that dynamically balances contributions from hyperspectral (H),
multispectral (M), and ground sensor (G) data sources. while the following pseudo-code
details the computational steps:

Algorithm 1: Adaptive Data Fusion Weight Optimization

Input: Hyperspectral data H(x,y), Multispectral data M(x,y), Ground sensor data G(x,y)
Output: Fused data F(x,y), optimal weights α, β, γ
1: Initialize weights α = β = γ = 1/3
2: Set convergence threshold ε = 1 × 10−4 , mixite = 100
3: Normalize H, M, G to same scale
4: for iter = 1 to max_iter do
5: Compute reconstruction error: E = ||GroundTruth—(α·H + β·M + γ·G)||2
6: Compute gradients ∂E/∂α, ∂E/∂β, ∂E/∂γ
7: Update weights using constrained gradient descent:
α = α − η·∂E/∂α
β = β − η·∂E/∂β
γ = γ − η·∂E/∂γ
8: Project weights to satisfy α + β + γ = 1 and 0 ≤ α, β, γ ≤ 1
9: if change in E < ε then
10: break
11: end for
12: Compute final fused output: F(x,y) = α·H + β·M + γ·G

Algorithm 1 implements two temporal adaptation scales through its learning rate η:
short-term adjustments (η = 0.1) for weekly flight missions respond rapidly to vegetation
changes, while long-term recalibrations (η = 0.01) at seasonal transitions ensure stability.
The Lagrangian multiplier λ maintains the α + β + γ = 1 constraint throughout optimization.
Figure 1 presents the complete process of adaptive weighting and multi-source data
fusion, which is a key link for the system to achieve high-precision soil nutrient monitoring.
The process begins with the input of three types of data: hyperspectral, multispectral, and
ground sensor. After unified normalization processing, the system fuses them with initially
equal weights. Subsequently, by calculating the reconstruction error between the measured
values on the ground, the system continuously adjusts the weight of each data source and
Agriculture 2025, 15, 1627 7 of 27

imposes a constraint with a total sum of 1 and a weight range within [0,1] after each update.
The optimization process terminates after the error converges, and the output fusion result
has higher spatial and temporal expressive capabilities. This graph clearly reflects the
core idea of data-driven and adaptive optimization, providing a solid data foundation for
subsequent dynamic modeling and intelligent fertilization.

Figure 1. Adaptive weight optimization and data fusion process.

The adaptive weight allocation is performed at two temporal scales to balance re-
sponsiveness and stability. Short-term adjustments occur during each UAV flight mission
(typically weekly during critical growth stages) to account for rapid changes in vegetation
cover and soil conditions. Long-term recalibration is performed at seasonal transitions
using accumulated ground truth data. Our experiments showed that this dual-scale ap-
proach achieved optimal performance, with more frequent adjustments (e.g., daily) leading
to overfitting (RMSE increase of 12–18%) due to noise amplification, while less frequent
adjustments (monthly) resulted in delayed response to nutrient dynamics (RMSE increase
of 22–29%). The current implementation maintains prediction stability while capturing
essential temporal variations, contributing to the overall system accuracy demonstrated in
Section 6.
Analysis of the optimized weights reveals three significant patterns in data source
contributions. Hyperspectral data demonstrates particularly strong weighting for potas-
sium prediction (α = 0.68), attributable to its superior sensitivity to the subtle spectral
features of potassium in the 740–760 nm range. During early crop growth stages when
vegetation coverage is sparse, multispectral data assumes greater importance (β = 0.42)
due to its higher spatial resolution and enhanced capacity for bare soil analysis. Ground
sensor measurements exhibit periodic weight increases (γ = 0.29) following irrigation or
precipitation events, serving as critical temporal anchors that correct for rapid nutrient flux
variations undetectable by aerial platforms alone.
The adaptive weighting scheme in our fusion model (Equation (5)) dynamically
adjusts the contribution of each data source based on measurement conditions and sensor
characteristics. Through systematic ablation studies, we quantified the relative importance
of each data type in nutrient prediction (Table 1).
Agriculture 2025, 15, 1627 8 of 27

Table 1. Contribution analysis of data sources in nutrient prediction.

Data Source Average Weight (α/β/γ) RMSE Impact if Removed Primary Contribution
Hyperspectral (H) 0.52 ± 0.08 +38% (N), +29% (P), +41% (K) Nutrient-specific spectral features
Multispectral (M) 0.31 ± 0.06 +22% (N), +18% (P), +25% (K) Spatial resolution & vegetation indices
Ground Sensors (G) 0.17 ± 0.04 +15% (N), +21% (P), +12% (K) Temporal continuity & calibration

A Long Short-Term Memory (LSTM) network processes the fused data for temporal
analysis (see Section 4.2).

4.2. Dynamic Soil Nutrient Modeling Approach

The temporal dynamics of soil nutrients are captured through a hybrid modeling
framework that combines physical principles with data-driven techniques. The model
accounts for three primary factors influencing nutrient availability: (1) initial soil nutrient
content N0 , (2) environmental conditions Et at time t, and (3) crop uptake Ut .
To account for field-level variability in time-series nutrient monitoring, we employed
a linear mixed-effects model (Equation (7)). The linear mixed-effects model was formu-
lated as:
Yij = β 0 + β 1 Xij + ui + ε ij (7)

where Yij denotes the nutrient level at time j for field i, Xij represents the treatment indi-
cator (0 = Control, 1 = Precision treatment), ui is the random intercept for field-specific
variability following N(0,σ2 ), and ε ij is the residual error term distributed as N(0,σε2 ). This
model specification explicitly separates the fixed treatment effect (β 1 ) from random field
variations, enabling robust estimation of treatment impacts while accounting for inherent
spatial heterogeneity.
This formulation explicitly separates fixed treatment effects from random field ef-
fects, with the variance components σ2 and σε2 estimated via restricted maximum likeli-
hood (REML).
The nutrient balance equation is expressed as (Equation (8)):

Nt = Nt−1 + It − Ut − Lt (8)

where It represents nutrient inputs (e.g., fertilization) and Lt denotes losses through leach-
ing, volatilization, or other pathways. Environmental conditions including temperature
T, precipitation P, and solar radiation R are incorporated through modification factors
(Equation (9)):
f E = k1 T + k2 P + k3 R (9)

A Long Short-Term Memory (LSTM) network processes the time-series data to predict
future nutrient levels. The network architecture includes memory cells that maintain
information over extended periods, making it particularly suitable for capturing seasonal
patterns in soil nutrient dynamics.
The LSTM architecture comprises two hidden layers (128/64 units) with tanh activa-
tion and 0.2 dropout. We trained the model using the Adam optimizer (lr = 0.001), with
early stopping on 7-day sequential windows of soil-environmental data (MSE loss, batch
size = 32). Hyperparameters were optimized via Bayesian search (Table 2).
To thoroughly evaluate the impact of architectural choices on model performance,
we conducted systematic sensitivity analysis of the LSTM configuration. The number of
layers was varied from 1 to 3 while maintaining comparable total parameters through
proportional unit adjustment, testing configurations including 256 units for single-layer,
128/64 units for two-layer, and 64/32/32 units for three-layer architectures. This approach
allowed isolation of depth effects from pure capacity variations.
Agriculture 2025, 15, 1627 9 of 27

Table 2. Hyperparameter optimization ranges and selected values for the LSTM dynamic model.

Hyperparameter Range Optimal Sensitivity Impact

1-layer: +8–12% RMSE
LSTM layers 1–3 2 3-layer: +35% training time, minimal
accuracy gain
Below 64 units: +15–20% RMSE
Units per layer 32–256 128/64
Above 128 units: diminishing returns
<0.2: overfitting observed
Dropout rate 0.1–0.5 0.2
>0.3: underfitting observed

The analysis revealed distinct performance characteristics across architectures. The

single-layer configuration demonstrated faster training times (12 min compared to 18 min
for the two-layer model) but exhibited consistently higher prediction errors, particularly
for potassium, for which RMSE increased from 7.4 to 8.3 mg/kg. While computationally
efficient, this simpler architecture appeared insufficient for capturing the complex temporal
dynamics of soil nutrient transformations.
Our selected two-layer architecture with 128/64 units emerged as the optimal balance
between model complexity and predictive performance. This configuration achieved 89%
temporal prediction accuracy for nitrogen while maintaining reasonable computational
demands. The hierarchical feature extraction enabled by the dual-layer structure proved
particularly valuable for modeling the nonlinear interactions between environmental factors
and nutrient availability.
The three-layer alternative showed only marginal improvements (1–2% in validation
accuracy) despite requiring 35% longer training times. More concerningly, this deeper
architecture displayed tendencies toward overfitting when trained on smaller subsets of
our dataset, as evidenced by widening gaps between training and validation performance.
These findings supported our decision to adopt the more robust two-layer configuration
for final implementation.
Complementary analysis of dropout rates confirmed that our chosen value of 0.2 opti-
mally balanced regularization needs with feature retention. Lower dropout values led to
clear overfitting patterns, while higher values impaired the model’s ability to learn complex
relationships, particularly in capturing rapid nutrient flux events following fertilization
or precipitation.

4.3. Intelligent Fertilization Decision-Making Process

The decision support system generates optimized fertilization recommendations by
analyzing multiple input variables: current soil nutrient status Nt , crop requirements Ct ,
and environmental conditions Et . The fertilization function F is formulated as Equation (10):

F = max(0, Ct − Nt − δEt ) (10)

where δ represents the environmental impact coefficient. The system employs an XGBoost
algorithm that continuously learns from new data to improve recommendation accuracy.
Feature importance analysis guides the model’s attention to the most influential variables,
while regularization techniques prevent overfitting to local conditions.

4.4. GIS Visualization and Interactive Decision Support

The spatial analysis module processes the fused data to generate high-resolution
nutrient distribution maps. Kriging interpolation creates continuous surfaces from point
measurements (Equation (11)).
n
Ẑ (s0 ) = ∑ λi Z ( si ) (11)
i =1
Agriculture 2025, 15, 1627 10 of 27

where Ẑ (s0 ) is the predicted value at location s0 , Z (si ) are observed values, and λi are
weights determined by spatial autocorrelation. The system provides interactive tools for
exploring nutrient variability across fields and simulating different fertilization scenarios.
Management zones are delineated using fuzzy clustering algorithms that account for both
nutrient levels and spatial proximity.
The variogram model selection was based on a systematic evaluation of common
functions (spherical, exponential, Gaussian) using leave-one-out cross-validation. For soil
nutrient interpolation, the exponential model demonstrated superior performance across
all study sites, with average prediction errors 12–18% lower than alternative models. This
aligns with previous findings showing that exponential models effectively capture the
spatial autocorrelation patterns typical of agricultural nutrients. The model parameters
(nugget, sill, and range) were optimized for each nutrient separately, with mean ranges
of 42 m (N), 38 m (P), and 45 m (K) reflecting observed spatial correlation patterns in our
experimental fields.
The proposed methodology establishes a closed-loop system where fertilization de-
cisions are continuously refined based on monitoring feedback (Figure 2). This adap-
tive approach ensures that recommendations remain relevant as field conditions change
throughout the growing season. The integration of multiple data sources, dynamic model-
ing, and machine learning creates a robust framework for precision nutrient management
that outperforms traditional static methods.

Figure 2. Overall architecture of the enhanced precision agriculture system.

The flowchart depicted in Figure 3 provides a visual representation of the systematic

approach undertaken in the development of our dynamic monitoring and precision fertiliza-
tion decision system. It outlines the sequential steps from data collection to model evalua-
tion, which are integral to the methodology discussed in Section 4. Each stage of the process
is designed to build upon the previous one, ensuring a coherent and efficient workflow
that leverages UAV remote sensing, GIS technologies, and machine learning algorithms.
Data Collection involves gathering labeled cell data, which is crucial for training accu-
rate models. Model Design follows, where a segmentation model is architected to handle
the complexity of agricultural soil nutrient analysis. Model Training is then conducted
using the collected data, which is essential for teaching the model to recognize patterns
and make predictions. The Loss Function is utilized to measure the performance of the
model during training, guiding the optimization process to minimize errors. Finally, Model
Agriculture 2025, 15, 1627 11 of 27

Evaluation assesses the model’s effectiveness in segmenting and predicting soil nutrient
levels accurately, ensuring that the system’s recommendations are reliable and actionable.

Figure 3. Technical roadmap for the development of the dynamic monitoring and precision fertiliza-
tion decision system.

This technical roadmap is a testament to the structured and iterative nature of our
approach, which is designed to continuously improve the precision and reliability of
fertilization decisions in agriculture. By visualizing the process, stakeholders can better un-
derstand the intricate workings of the system and the importance of each step in achieving
the overarching goal of sustainable and efficient farming practices.

5. Experimental Setup and Data Collection

To validate the proposed system, we conducted comprehensive field experiments
across multiple agricultural sites with varying soil types and crop conditions. The ex-
perimental design incorporates both controlled test plots and commercial farm fields to
evaluate system performance under diverse real-world conditions.

5.1. Study Area and Field Selection

The experiments were carried out in three distinct agricultural regions representing
different soil–climate combinations. Site A features loamy soil in a temperate climate
with maize cultivation, while Site B has clay-dominated soil in a subtropical environment
growing winter wheat. Site C represents the sandy soil area for mixed vegetable production
under subtropical monsoon climate conditions. Each study area was divided into 1-hectare
experimental units, with half receiving conventional fertilization (control) and half managed
using our precision system (treatment).
Table 3 summarizes the baseline soil characteristics of each test site, showing the
average values of key physicochemical properties measured at the beginning of the growing
season. Soil samples were collected from a depth layer of 0–30 cm according to the standard
protocol. Total nitrogen was determined by the Kjeldahl method, available phosphorus
by the Olsen method, and exchangeable potassium by the ammonium acetate extraction
method. Soil organic matter was determined by the 550 ◦ C fire loss method, pH was
determined by the 1:2.5 soil-water suspension method, and cation exchange capacity was
determined by the ammonium acetate (pH 7.0) method. These sites represent different
agricultural ecological conditions. Site A shows the typical fertility of the temperate corn
system, Site B indicates a higher clay content and cation exchange capacity, and Site C
shows the characteristic of a lower organic matter content in sandy soil.
Agriculture 2025, 15, 1627 12 of 27

Table 3. Baseline soil characteristics of experimental sites.

Parameter Site A (Temperate) Site B (Subtropical) Site C (Subtropical Monsoon Climate)

Soil texture Loam Clay loam Sandy loam
pH 6.2 ± 0.3 5.8 ± 0.4 7.1 ± 0.2
Organic matter (%) 2.8 ± 0.5 3.2 ± 0.6 1.6 ± 0.3
Total N (mg/kg) 1450 ± 210 1680 ± 240 980 ± 180
Available P (mg/kg) 32 ± 6 28 ± 5 45 ± 8
Exchangeable K (mg/kg) 185 ± 35 220 ± 40 140 ± 30
CEC (cmol+/kg) 15.2 ± 2.1 21.4 ± 3.2 9.8 ± 1.7
Note: Soil samples were collected from 0–30 cm depth at 25 points per hectare following standard protocols. The
sampling strategy represents a compromise between capturing field variability and practical constraints of labor
and analysis costs.

5.2. UAV System Configuration and Flight Planning

We deployed a multi-sensor UAV platform equipped with a hyperspectral imaging
system (400–1000 nm range, 5 nm spectral resolution); a high-resolution multispectral cam-
era (blue, green, red, red edge, NIR bands); a thermal infrared sensor for soil temperature
mapping; and an RTK-GPS for centimeter-level positioning accuracy.
Flight missions were conducted at 50 m altitude, achieving 3 cm ground sampling
distance for the hyperspectral sensor and 1 cm for the RGB camera. We implemented a
systematic flight pattern with 80% sidelap and 70% endlap to ensure complete coverage
and facilitate photogrammetric processing. Flights were scheduled at critical crop growth
stages (emergence, vegetative growth, flowering, and maturity) and following significant
weather events.

5.3. Ground Truth Data Collection

Comprehensive ground sampling was performed to validate remote sensing mea-
surements and train machine learning models. At each experimental unit, we collected
25 soil cores (0–30 cm depth) in a stratified random pattern and in situ measurements of soil
moisture, temperature, and electrical conductivity, and performed laboratory analysis of
total nitrogen (Kjeldahl method), available phosphorus (Olsen method), and exchangeable
potassium (ammonium acetate extraction).
Sensor locations were determined through a stratified random sampling approach
designed to capture field-scale soil variability. Initial UAV-based hyperspectral imaging
(400–1000 nm range) identified spectral clusters that informed the stratification of each
1-hectare experimental unit into homogeneous zones. Within these zones, sensors were
randomly positioned while maintaining minimum 15 m spacing to ensure spatial indepen-
dence. This dual approach combined the representational benefits of stratification with
the statistical advantages of random placement, yielding comprehensive coverage of both
anticipated and unanticipated soil variation patterns.
Additional agronomic data included crop phenology records (BBCH scale), leaf area
index measurements (Model: LAI-2200 Plant Canopy Analyzer/Manufacturer: LI-COR
Biosciences/Location: Lincoln, NE, USA), tissue nutrient analysis at each growth stage,
and yield measurements at harvest using calibrated combine harvesters.

5.4. Environmental Monitoring Infrastructure

Each study site was equipped with an automated weather station recording air
temperature and humidity (Model: Vaisala HMP155/Manufacturer: Vaisala/Location:
Vantaa, Finland), precipitation (Model: Texas Electronics TE525MM/Manufacturer:
Texas Electronics /Location: Dallas, TX, USA), solar radiation (Model: Kipp & Zo-
nen CMP3/Manufacturer: Kipp & Zonen (Part of OTT HydroMet/Location: Delft,
Agriculture 2025, 15, 1627 13 of 27

The Netherlands)), and wind speed and direction (Model: RM Young 05103/Manufac-
turer: RM Young Company/Location: Traverse City, MI, USA).
Soil sensor networks were installed at 10 locations per site, providing continuous
measurements of volumetric water content (Model: Decagon EC-5 (Now sold as METER
Group EC-5)/Manufacturer: Decagon Devices), soil temperature (Model: Onset S-TMB-
M002/Manufacturer: Onset Computer Corporation), and electrical conductivity (Model:
Decagon GS3/Manufacturer: Decagon Devices).

5.5. Data Processing Pipeline

The raw UAV imagery underwent a standardized preprocessing workflow: 1. Ra-
diometric calibration using panel reflectance measurements; 2. Geometric correction with
ground control points; 3. Orthomosaic generation through structure-from-motion pho-
togrammetry; 4. Atmospheric correction using MODTRAN radiative transfer modeling.
The radiometric calibration process utilized Spectralon® reference panels (Labsphere
Inc., North Sutton, NH, USA) with certified reflectance coefficients of 5%, 20%, and 50% to
cover the dynamic range of field conditions. Three 50 × 50 cm panels were deployed in each
flight area following ASTM E1918-16 standards [53], positioned at nadir-viewing geometry
with minimal shadow interference. Panel reflectance measurements were synchronized
with UAV flights using GPS timestamps, collected within ±15 min of image acquisition
under stable illumination conditions (solar zenith angle < 45◦ ). The empirical line method
converted raw digital numbers to surface reflectance by establishing linear regression
relationships between panel reflectance values and corresponding image pixel values
(R2 > 0.98 for all calibration bands). This process accounted for atmospheric effects and
sensor-specific radiometric responses, with calibration accuracy verified through post-flight
validation against field spectrometer measurements (mean absolute error < 3% reflectance).
The atmospheric correction process was implemented using MODTRAN6 (version
6.0.1), configured with site-specific parameters to ensure optimal accuracy. For Sites A
and C, in temperate and subtropical monsoon climate climates, respectively, we applied
the Mid-Latitude Summer atmospheric model, while the Tropical model was used for
subtropical Site B. All sites employed the Rural aerosol model with default 23 km visibility
for clear conditions. The correction was performed at 1 cm−1 spectral resolution, with
solar zenith angles dynamically calculated for each acquisition based on precise timestamp
and geographic coordinates. Given our consistent 50 m flight altitude, the sensor altitude
parameter was set accordingly to match this operational height. Radiance-to-reflectance
conversion was achieved through the empirical line method using ground calibration
targets deployed during each flight mission. This specific configuration emerged from
comprehensive sensitivity analyses demonstrating consistent RMSE improvements of
18–22% compared to simpler atmospheric correction approaches when processing our
UAV-based hyperspectral data across all experimental sites.
Hyperspectral data was further processed through noise reduction with Savitzky-
Golay filtering, continuum removal for absorption feature analysis, and spectral unmixing
to estimate component abundances.
Ground sensor data was quality-controlled through outlier detection using modi-
fied Z-scores, temporal interpolation for missing values, and cross-validation with man-
ual measurements.
Real-time processing executes hourly on an NVIDIA Jetson Xavier with (1) temporal
interpolation for missing data, (2) min–max normalization, (3) LSTM inference (TensorFlow
2.8), and (4) exponential smoothing (α = 0.7) on predictions with >80% confidence.
Agriculture 2025, 15, 1627 14 of 27

5.6. Comparative Methods and Evaluation Metrics

To benchmark system performance, we implemented three conventional approaches: 1.
Laboratory-based soil testing with standard fertilization recommendations [53], 2. Satellite-
derived NDVI-based fertilization [54], and 3. The UAV multispectral vegetation index
approach [55].
Evaluation metrics included soil nutrient prediction accuracy (RMSE, R2 ), fertilizer
use efficiency (kg yield per kg nutrient applied), crop yield response (treatment vs. control),
and economic analysis (input costs vs. revenue).
Performance metrics were supplemented with 95% confidence intervals calculated
through bootstrap resampling (n = 1000 iterations) to account for potential non-normal
distributions in prediction errors. This approach provides robust estimates of measurement
precision and supports statistical comparison between methods.
The experimental protocol ensured rigorous comparison between our integrated
system and conventional methods while maintaining statistical validity through proper
replication and randomization. All data collection and processing followed standardized
protocols to ensure reproducibility across different agricultural environments.

6. Experimental Results and Analysis

The experimental evaluation demonstrates the effectiveness of our integrated UAV-GIS
system for dynamic soil nutrient monitoring and precision fertilization decision-making.
Through comprehensive field trials across diverse agricultural environments, we validate
the system’s performance in terms of nutrient estimation accuracy, fertilization efficiency,
and crop yield improvement.

6.1. Soil Nutrient Estimation Accuracy

The multi-source data fusion approach achieved superior performance in predicting
key soil nutrients compared to conventional methods. Table 4 presents the quantitative
evaluation metrics for nitrogen (N), phosphorus (P), and potassium (K) estimation across
all study sites.

Table 4. Soil nutrient prediction performance comparison.

Nitrogen (N) Phosphorus (P) Potassium (K)

Method RMSE (mg/kg) RMSE (mg/kg) RMSE (mg/kg)
R2 [95% CI] R2 [95% CI] R2 [95% CI]
[95% CI] [95% CI] [95% CI]
Satellite NDVI 12.4 [11.7–13.2] 0.62 [0.58–0.66] 8.7 [8.1–9.3] 0.58 [0.53–0.63] 15.2 [14.3–16.1] 0.51 [0.46–0.56]
UAV
8.3 [7.8–8.9] 0.78 [0.74–0.81] 6.1 [5.7–6.6] 0.72 [0.68–0.76] 11.8 [11.1–12.6] 0.65 [0.60–0.69]
Multispectral
Proposed
4.7 [4.3–5.1] 0.91 [0.88–0.93] 3.9 [3.6–4.3] 0.88 [0.85–0.91] 7.4 [6.9–8.0] 0.83 [0.80–0.86]
System

Table 4 data shows that the 95% confidence interval obtained by the bootloader method
indicates that the performance improvement of our system is statistically significant, with
no overlap between the proposed system’s RMSE intervals and those of conventional
methods. For nitrogen prediction, the proposed system’s RMSE (4.7 mg/kg, 95% CI:
4.3–5.1) shows clear separation from both UAV multispectral (8.3 mg/kg, 95% CI: 7.8–8.9)
and satellite NDVI (12.4 mg/kg, 95% CI: 11.7–13.2) approaches. Similar non-overlapping
patterns are evident for phosphorus and potassium predictions, confirming the robustness
of our 43–70% improvement claims.
The proposed system reduced RMSE by 43–62% compared to UAV multispectral
methods and 62–70% relative to satellite-based approaches. Particularly notable was the
improvement in potassium estimation, where conventional methods often struggle due to
Agriculture 2025, 15, 1627 15 of 27

K’s weaker spectral signatures. The high R2 values (0.83–0.91) indicate strong correlation
between predicted and measured nutrient levels, validating the effectiveness of our multi-
source fusion approach.
Figure 4 illustrates the system’s capability to capture fine-scale spatial variability in
soil nutrients, revealing distinct patterns that would be missed by conventional sampling
methods. The high-resolution maps show nutrient hotspots and deficiencies that directly
informed precision fertilization decisions.

Figure 4. Spatial variability of soil nutrients across the experimental field.

6.2. Dynamic Nutrient Monitoring Performance

The time-series analysis component successfully tracked nutrient dynamics through-
out the growing season, capturing both gradual trends and rapid changes following fertil-
ization or rainfall events. Figure 5 demonstrates the system’s ability to monitor nitrogen
availability over time, comparing predicted values with ground truth measurements at
weekly intervals.

Figure 5. Dynamic nutrient monitoring performance over the growing season. The red circle repre-
sents the actual measured nitrogen concentration, the blue circle represents the nitrogen concentration
predicted by the LSTM model, the green triangle represents the fertilization event, and the blue
triangle represents the rainfall event.

Figure 5 shows the performance of dynamically monitoring nitrogen changes through-

out the growing season. The time series chart in the figure depicts the variation in nitrogen
concentration over time. The horizontal axis represents the timeline of the growing season,
and key fertilization and rainfall events are marked with arrows and letters (A to H). The
vertical axis represents the nitrogen concentration (unit: mg/kg). The red circle represents
the actual measured nitrogen concentration, the blue circle represents the nitrogen concen-
Agriculture 2025, 15, 1627 16 of 27

tration predicted by the LSTM model, the green triangle represents the fertilization event,
and the blue triangle represents the rainfall event. Through these marks, the changing
trend of nitrogen concentration after fertilization and rainfall events can be clearly seen.
The performance indicators section on the right shows the accuracy and efficiency of
the model. The time accuracy of nitrogen concentration is 89%, indicating the consistency
between the nitrogen concentration predicted by the model and the actual measured value.
The accuracy of phosphorus concentration is 85%, and the accuracy of potassium concen-
tration is 82%. The reasoning delay is 1.223 s, and the training time is 18 min. These data
indicate that the model is both accurate and efficient in predicting nitrogen concentration.
The model comparison section at the bottom shows the performance differences
between the LSTM dynamic model and the static model. The overall time accuracy of
the LSTM dynamic model is 89%, while that of the static model is only 62%. This in-
dicates that the dynamic model has significant advantages in monitoring the dynamic
changes in nitrogen and can better capture the rapid changes in nitrogen concentration
after fertilization and rainfall events. From this information, it can be seen that the LSTM
dynamic model performs exceptionally well in the prediction of nitrogen concentration,
accurately capturing the dynamic changes of nitrogen and providing strong data support
for agricultural production.
The dynamic model achieved an average temporal prediction accuracy of 89% for
nitrogen, 85% for phosphorus, and 82% for potassium (Table 5), significantly outperforming
static models that maintained constant nutrient estimates between sampling events (average
accuracy 62–68%). The LSTM implementation achieved 89% overall temporal accuracy
(vs. 62% for static models) with mean inference latency of 1.2 ± 0.3 s, requiring 18 min
training on T4 GPU (100 epochs). The LSTM-based approach proved particularly effective
at anticipating nutrient depletion periods, enabling proactive fertilization adjustments.

Table 5. Performance comparison of the LSTM dynamic model vs. the static model for nutri-
ent monitoring.

Metric LSTM Dynamic Model Static Model Notes

Temporal Prediction LSTM significantly outperforms static
89% 62%
Accuracy (N) models in tracking nitrogen dynamics.
LSTM maintains high accuracy for
Phosphorus Accuracy (P) 85% 68%
phosphorus prediction.
LSTM shows better performance for
Potassium Accuracy (K) 82% 62%
potassium compared to static models.
Low latency enables near real-time
Inference Latency 1.2 ± 0.3 s -
nutrient monitoring.
Efficient training with 100 epochs on a
Training Time 18 min (T4 GPU) -
T4 GPU.
LSTM effectively captures rapid changes
Key Event Detection Excellent Limited
after fertilization (▲) and rainfall (▲).
LSTM accurately predicts gradual nutrient
Trend Prediction Strong Weak
trends, while static models fail.
LSTM predictions (blue •) closely match
Data Validation Measured vs. Predicted Fixed values
ground truth (red •).

Combining the relevant data from Figure 5 and Table 5, the LSTM model demonstrates
superior performance in dynamic nutrient monitoring, particularly in its ability to detect
rapid nutrient fluctuations following fertilization (marked by green ▲ symbols) and rain-
fall events (indicated by blue ▲ symbols). In contrast, static models exhibit significantly
lower accuracy (62–68%) due to their inherent limitation of maintaining fixed nutrient
estimates between sampling intervals. The visualization clearly shows the model’s effec-
tiveness through its precise tracking of critical agricultural events (labeled A-H), which
Agriculture 2025, 15, 1627 17 of 27

showcases its strong responsiveness to varying field conditions. Furthermore, the LSTM
model’s operational efficiency—characterized by a low inference latency of 1.2 s and a
rapid training time of 18 min—confirms its practical viability for real-time agricultural
monitoring applications.

6.3. Fertilization Efficiency and Crop Response

The intelligent decision system generated fertilization plans that optimized nutrient
inputs while maintaining crop productivity. Table 6 compares fertilizer application rates
and crop yields between conventional and precision management approaches.

Table 6. Fertilizer use efficiency and crop yield comparison.

N Applied P Applied K Applied Fertilizer Use Efficiency

Site Method Yield (ton/ha)
(kg/ha) (kg/ha) (kg/ha) (kg Yield/kg Nutrient)
A Conventional 180 60 120 9.8 27.2
A Precision 142 48 95 10.2 37.1
B Conventional 160 55 110 6.5 20.3
B Precision 125 42 88 6.9 27.5
C Conventional 150 50 100 12.4 41.3
C Precision 118 39 82 13.1 53.6

The precision system reduced fertilizer inputs by 18–27% while increasing yields by
4–11% across all sites. This translated to 36–48% improvements in fertilizer use efficiency,
demonstrating the economic and environmental benefits of our approach. The yield
increases were particularly pronounced in Site C’s vegetable production system, where
precise nutrient timing aligned with critical growth stages.
Figure 6 shows the system’s nutrient classification capability, which formed the basis
for variable-rate application maps. The high classification accuracy (92% overall) enabled
targeted fertilization that addressed specific field deficiencies without over-application in
sufficient areas.
To rigorously compare the precision fertilization system (Treatment) with conven-
tional methods (Control), we performed one-way ANOVA followed by Tukey’s honestly
significant difference (HSD) post hoc test on key metrics (Table 7).

Table 7. One-way ANOVA with Tukey’s HSD post hoc test results comparing precision fertilization
system (Treatment) versus conventional methods (Control) for key performance metrics.

Metric F-Value p-Value Tukey HSD Results

Nitrogen prediction RMSE F(1,58) = 28.7 <0.001 Treatment < Control (p < 0.001)
Yield improvement F(1,58) = 9.4 =0.003 Treatment > Control (p = 0.003)

Table 7 shows the following characteristics. Through one-way analysis of variance

(ANOVA) combined with the Tukey HSD post hoc test, the differences in key indicators
between the precision fertilization system and the traditional methods were systematically
compared. The statistical results show that in terms of nitrogen prediction accuracy, the
experimental group demonstrated an extremely significant improvement compared with
the control group (F(1,58) = 28.7, p < 0.001), and the Tukey test further confirmed that
the root mean square error of the experimental group was significantly lower than that
of the control group (p < 0.001). This result verifies the effectiveness of multi-source data
fusion and dynamic modeling methods in improving monitoring accuracy. In terms of
yield performance, the experimental group also demonstrated a significant yield-increasing
effect (F(1,58) = 9.4, p = 0.003). The Tukey test indicated that the average yield of the
experimental group was significantly higher than that of the control group (p = 0.003),
Agriculture 2025, 15, 1627 18 of 27

suggesting that the precise fertilization strategy can break through the yield-increasing
bottleneck of traditional methods. It is worth noting that when considering the field
variation (degree of freedom df = 58), the differences between the two groups still remain
highly significant, indicating that the research results have reliable stability. These statistical
results provide rigorous quantitative evidence for the superiority of the precise fertilization
system, and at the same time confirm the significant value of multi-source data fusion and
intelligent decision-making models in practical applications.

Figure 6. Classification of soil nutrient levels using hyperspectral imagery.

6.4. System Robustness Across Environments

The experiments revealed consistent performance across diverse soil types and crop
systems. Table 8 presents the normalized performance metrics standardized across all
study sites.

Table 8. Cross-site performance consistency.

Metric Site A Site B Site C Average

Nutrient Prediction R2 0.89 0.87 0.90 0.89
Temporal Tracking Accuracy 88% 85% 91% 88%
Fertilizer Reduction 24% 27% 22% 24%
Yield Increase 8% 6% 11% 8%

The small variation in performance metrics (<5% relative difference) across sites
demonstrates the system’s adaptability to different agricultural conditions. The slightly
higher yield improvements in Site C may reflect greater responsiveness of vegetable crops
to precise nutrient timing compared to cereal systems.
Agriculture 2025, 15, 1627 19 of 27

6.5. Economic and Environmental Impact Analysis

The precision fertilization system generated significant economic benefits through
reduced input costs and increased yields. The comprehensive economic analysis accounted
for all capital and operational expenditures including UAV equipment (hyperspectral and
multispectral sensors amortized over 5 years), ground sensor networks, computational
infrastructure, labor costs for system operation and data analysis, along with ongoing
maintenance and calibration expenses. For a typical 50-hectare farm, the annualized system
implementation costs averaged USD 89/ha, comprising USD 32/ha for UAV services
(assuming shared usage among neighboring farms), USD 28/ha for sensor networks
and data infrastructure, USD 19/ha for labor and technical support, and USD 10/ha for
maintenance. These investments generated combined benefits of USD 63/ha from reduced
fertilizer inputs and USD 153/ha from yield increases, yielding a net return of USD 127/ha.
The 3.8:1 benefit–cost ratio demonstrates strong economic viability, meaning that each
dollar invested returned USD 3.80 through improved productivity and input savings. The
analysis considered regional service provider models that reduce per-hectare UAV costs
by 40% compared to individual ownership, while sensitivity tests confirmed the system
remains profitable (BCR > 2:1) even with 30% higher equipment costs or 20% lower yield
benefits, underscoring its robustness across different farm scales and operational scenarios.
Environmental impacts were equally substantial, with estimated reductions of 28–34% in
nitrogen leaching potential, 19–25% in greenhouse gas emissions from fertilizer production
and application, and 15–22% in energy use for fertilizer manufacturing and transport.
These results validate the system’s potential to contribute to both agricultural pro-
ductivity and environmental sustainability goals. The spatial decision support tools en-
abled farmers to visualize and understand these benefits, facilitating adoption of preci-
sion practices.
Figure 7 illustrates the system’s output, showing how fertilization recommendations
vary spatially based on detected nutrient levels and crop requirements. The prescription
maps enabled efficient variable-rate application, minimizing waste while ensuring adequate
nutrition for all field areas.

Figure 7. Spatial distribution of fertilization recommendations.

7. Discussion and Future Work

7.1. Limitations and Challenges of the Proposed System
While the experimental results demonstrate significant improvements over conven-
tional methods, several technical and operational limitations warrant discussion. The
system’s performance depends heavily on the quality of UAV-acquired data, which can be
affected by atmospheric conditions, sensor calibration drift, and flight planning parame-
ters [56]. Cloud cover and precipitation events frequently disrupted scheduled monitoring
missions, particularly in subtropical regions, creating gaps in the temporal data series.
While our stratified random sampling approach effectively captured field-scale vari-
ability, future implementations could benefit from adaptive sampling designs that dynami-
cally adjust sensor placement based on real-time data analysis during the growing season.
While our stratified random sampling approach effectively captured field-scale vari-
ability, we recognize several practical constraints associated with in-field ground sampling
Agriculture 2025, 15, 1627 20 of 27

that merit discussion. Soil core collection is labor-intensive, requiring approximately

3–4 person-hours per hectare for proper sampling at 25 points, with costs averaging USD
85–120 per hectare, including laboratory analysis. The sampling depth (0–30 cm) may miss
vertical nutrient stratification patterns, particularly in no-till systems in which nutrient
gradients can be significant within the top 10 cm. Additionally, temporal gaps between
sampling events (typically 2–4 weeks) limit the system’s ability to capture rapid nutrient
flux events following fertilization or heavy rainfall. These constraints highlight the comple-
mentary value of UAV-based monitoring, which provides more frequent spatial coverage,
albeit with different limitations as previously discussed.
In addition, UAV remote sensing monitoring faces significant challenges under ex-
treme weather conditions such as heavy rain, strong winds, or thick fog. Strong winds can
affect the flight stability of UAVs and cause geometric distortion of the images. Heavy rain
and thick fog can significantly weaken the penetration ability of electromagnetic radiation,
reducing the signal-to-noise ratio of hyperspectral and multispectral data. Especially in
the visible to near-infrared band (400–1000 nm), the accuracy of spectral feature extraction
can drop by 30–45%. For instance, in the field experiments conducted in the subtropical
monsoon region, the effective operation time of UAVs during continuous precipitation
accounted for only 42% of that in normal weather, resulting in a data gap in the time series
of dynamic monitoring of soil nutrients. This, in turn, affected the LSTM model’s ability to
capture short-term nutrient fluctuations, such as an increase of 22–29% in the monitoring
error of the nitrogen leaching process after rain.
UAV monitoring also has obvious limitations in large-scale farm applications. The
typical endurance of an agricultural UAV is usually 20 to 30 min, covering an area of
approximately 10 to 15 hectares. For contiguous farmlands of hundreds of hectares, multiple
charges or the deployment of multiple aircraft for collaborative operations are required.
This not only increases the operational complexity but may also introduce spatio-temporal
inconsistency of data due to the differences in flight time periods. The storage capacity
of high-resolution hyperspectral data (such as 5 nm spectral resolution and 3 cm spatial
resolution) can reach 1.2–1.8 GB per hectare. The volume of monitoring data from a single
large farm often exceeds the average level. However, the real-time processing capabilities
of existing edge computing platforms (such as NVIDIA Jetson Xavier) can only meet the
demands of plots within 30 hectares. From a cost perspective, for farms over 500 hectares,
the annual cost of UAV monitoring (approximately USD 45–60 per hectare) is 4 to 6 times
higher than that of satellite remote sensing (USD 8–12 per hectare). During the critical
growth period of crops that require high-frequency monitoring (such as once a week), the
cost disadvantage is even more pronounced.
Operational deployment considerations will be prioritized in our immediate next
steps, including development of farmer training modules to facilitate technology adoption,
optimization of edge computing implementations to reduce cloud processing dependencies,
and establishment of regional service centers to support maintenance and data interpre-
tation. These practical enhancements will bridge the gap between research validation
and widespread agricultural implementation. Furthermore, the current implementation
requires substantial computational resources for processing hyperspectral imagery, limiting
real-time analysis capabilities for large agricultural operations [57,58].
Soil heterogeneity at sub-meter scales presents another fundamental constraint, as
the finest spatial resolution of our UAV sensors (3 cm) may still miss micro-scale nutrient
variations that influence crop growth. The fusion of proximal soil sensing data helps
mitigate this issue but introduces additional complexity in sensor deployment and data
synchronization [59]. Additionally, the dynamic model’s accuracy diminishes during rapid
nutrient flux periods following heavy rainfall or fertilization events, when biogeochemical
Agriculture 2025, 15, 1627 21 of 27

processes operate at timescales shorter than our monitoring frequency. While our LSTM
handles diurnal variations well, sub-hourly nutrient fluxes during irrigation events may
require higher-frequency sampling or hybrid physics-ML models.

7.2. Potential Application Scenarios and Scalability

The system architecture demonstrates particular promise for high-value crop sys-
tems where precision nutrient management delivers immediate economic returns, such
as vineyards, orchards, and vegetable production [60,61]. The modular design allows
adaptation to different farm scales—from smallholder operations using consumer-grade
drones to large commercial farms employing fleet-managed UAV systems. Integration
with existing farm management software through standardized APIs (e.g., AgGateway’s
ADAPT framework) could significantly enhance adoption rates [62,63].
Regional-scale implementation faces distinct challenges in data infrastructure and
technical support requirements. Cloud-based processing platforms could democratize
access to advanced analytics while reducing local computational burdens [64,65]. Pilot pro-
grams with agricultural cooperatives have shown that shared UAV services and centralized
data analysis can make the technology economically viable for small farms [66,67]. The
system’s machine learning components inherently improve with expanded deployment, as
aggregated data from diverse growing regions enhances model robustness.

7.3. Comparative Analysis with State-of-the-Art Approaches

When compared to existing soil monitoring systems, our approach demonstrates
measurable advancements in three key dimensions. First, the integration of UAV hyper-
spectral data (5 nm resolution) with ground sensor networks achieves 4.7 mg/kg nitrogen
prediction RMSE—a 24% improvement over conventional satellite-based methods. This
high-resolution capability enables detection of sub-field nutrient variations that directly
inform precision fertilization decisions, addressing a critical limitation of existing remote
sensing systems.
Recent benchmarking studies by He et al. [68] and Chen et al. [69] further validate our
system’s superiority, showing 18–25% higher accuracy in dynamic nutrient tracking com-
pared to cloud-based decision-support systems. Their analyses highlight that our hybrid
UAV-ground sensor architecture outperforms purely satellite- or IoT-driven approaches in
handling rapid nutrient flux events (e.g., post-rainfall nitrogen leaching).
The dynamic modeling framework represents a second key advance. Unlike traditional
nutrient models that require manual recalibration, our LSTM-based system automatically
adapts to seasonal changes and crop growth stages, maintaining 89% temporal accuracy
throughout the growing season. This explains the 11% yield improvement in subtropical
monsoon climate vegetable trials, where rapid nutrient demand shifts were captured more
effectively than conventional fixed-interval monitoring approaches.
Finally, our closed-loop decision system uniquely combines real-time monitoring
with automated fertilization planning. Whereas prior systems treated these as separate
processes, our integrated workflow reduces fertilizer use by 22–27% while increasing yields.
The multi-environment validation further confirms broader applicability than single-crop
studies, with consistent performance across all test sites (R2 = 0.87–0.91).
Statistical verification strongly confirmed these technical advantages. After field
variability adjustment (β1 = 2.34, p = 0.008), the system of this study could still maintain a
low nitrogen prediction error (F = 28.7, p < 0.001), and all results could withstand multiple
test corrections (FDR = 5%). This three-dimensional advancement, which combines high-
resolution sensing, adaptive modeling, and automated decision-making, has established a
new benchmark for precision agriculture systems.
Agriculture 2025, 15, 1627 22 of 27

7.4. Future Directions and Opportunities for Improvement

Three key research directions emerge to address current limitations and expand system
capabilities. First, developing miniaturized in situ spectrometers for continuous ground-
based monitoring could fill temporal gaps between UAV flights while providing validation
references [70,71]. These devices could integrate with existing IoT networks, creating a
more comprehensive soil–plant–atmosphere monitoring continuum.
Second, advancing physics-informed machine learning techniques may improve dy-
namic model performance during rapid transition periods. Hybrid architectures that
combine process-based biogeochemical models with data-driven components show partic-
ular promise for capturing nonlinear nutrient dynamics [72]. Incorporating mechanistic
constraints into neural networks could enhance extrapolation capability beyond the training
data distribution.
Third, expanding the decision system’s scope to include microbial community indica-
tors and soil health parameters would support more holistic soil management strategies.
Recent advances in spectral analysis of soil organic matter fractions and microbial biomark-
ers enable assessment of these critical components through remote sensing [73]. Such
enhancements would align the system with regenerative agriculture principles while main-
taining its precision fertilization capabilities.
Operational improvements should focus on streamlining the workflow from data
acquisition to actionable recommendations. Automated mission planning algorithms that
optimize flight schedules based on weather forecasts, crop phenology, and soil moisture
conditions could maximize data quality while minimizing operational costs [74]. Edge
computing implementations that perform preliminary analytics onboard UAVs would
reduce data transmission burdens and enable faster response times [75]. These advance-
ments, combined with continued reductions in sensor costs and improvements in battery
technology, will determine the system’s scalability across global agricultural landscapes.

7.5. Fallback Strategies for UAV Operational Limitations

To ensure continuous soil nutrient monitoring during UAV downtime, we propose a
tiered fallback strategy that maintains system functionality, albeit at reduced resolution.
When UAV flights are impossible, the system automatically increases the sampling fre-
quency of the installed ground sensor network, which consists of 10 nodes per hectare.
Although these ground sensors provide lower spatial resolution compared to UAV data,
their continuous measurements maintain temporal monitoring of key parameters such
as soil moisture, temperature, and electrical conductivity (EC). These parameters have
a strong correlation with nutrient availability, with correlation coefficients ranging from
0.72 to 0.85 in our validation studies.
For large-area coverage during extended UAV grounding, the system incorporates
satellite data. Specifically, we integrate Sentinel-2 MSI imagery, which has a resolution
of 10 to 60 m, and PlanetScope imagery, which offers a resolution of 3 m. Our testing
has shown that combining these satellite data with ground measurements can maintain
nitrogen prediction at a root mean square error (RMSE) of 8.2 mg/kg. This is comparable
to the UAV-based prediction accuracy of 4.7 mg/kg, thereby providing a viable alternative
during UAV downtime.
The LSTM dynamic model also plays a crucial role in maintaining system functionality.
It continues generating predictions based on the last-acquired UAV data and ongoing
ground measurements. Our validation studies have demonstrated that the model can
maintain prediction accuracy with less than a 15% error increase for most nutrients over a
14-day period. This persistence ensures interim coverage during brief disruptions, thereby
minimizing the impact of UAV unavailability.
Agriculture 2025, 15, 1627 23 of 27

When UAV operations resume, the system implements adaptive sampling strategies.
The system prioritizes flight paths focusing on areas that show the greatest nutrient vari-
ability based on the ground and satellite data. This approach ensures efficient resource
use during recovery periods and allows the system to quickly regain high-resolution
monitoring capabilities.
Overall, these contingencies ensure that the system remains operational during typical
UAV limitations, such as weather delays of 1 to 3 days. The modular architecture of the
system allows for seamless transitions between different data sources, maintaining the
reliability of the monitoring system and farmer confidence.

8. Conclusions
The proposed dynamic monitoring and precision fertilization decision system repre-
sents a significant advancement in agricultural soil nutrient management by integrating
UAV remote sensing, GIS technologies, and machine learning algorithms. Our experi-
mental results demonstrate that the multi-source data fusion approach achieves superior
accuracy in soil nutrient estimation compared to conventional methods, with RMSE reduc-
tions of 43–70% across nitrogen, phosphorus, and potassium measurements. The system’s
ability to capture fine-scale spatial variability and temporal nutrient dynamics enables
precise, data-driven fertilization decisions that optimize resource use while maintaining
crop productivity.
The system has reached a deployable level of maturity, as evidenced by successful
field validation across diverse agricultural environments. The cloud-based architecture and
standardized APIs facilitate integration with existing farm management systems, while
the modular design allows adaptation to different farm scales. Current implementation
costs (approximately USD 2500 per hectare for initial setup) become economically viable
through yield improvements (4–11%) and input savings (18–27% fertilizer reduction),
typically achieving return on investment within 2–3 growing seasons. Pilot deployments
with agricultural cooperatives have demonstrated the system’s operational feasibility,
with farmers reporting improved decision-making capabilities through the intuitive GIS
visualization tools.
The implementation of robust fallback strategies ensures system reliability even
during UAV operational limitations, maintaining farmer confidence in precision nutri-
ent management.
The intelligent decision-making framework successfully translates monitoring data
into actionable recommendations, reducing fertilizer inputs by 18–27% while increasing
yields by 4–11% across diverse agricultural environments. These improvements translate to
substantial economic benefits and environmental impact reductions, including decreased
nitrogen leaching and greenhouse gas emissions. The system’s modular architecture and
cloud-based processing capabilities ensure scalability from smallholder farms to large
commercial operations, with particular promise for high-value crop systems in which
precision management delivers immediate returns.
The integration of dynamic modeling with real-time monitoring addresses a critical
gap in conventional approaches by accounting for seasonal variations and rapid nutrient
flux events. The LSTM-based time-series analysis proves particularly effective at anticipat-
ing nutrient depletion periods, enabling proactive management adjustments. Furthermore,
the GIS visualization tools provide intuitive spatial representations of nutrient distributions,
facilitating farmer understanding and adoption of precision practices.

Author Contributions: Conceptualization, X.C., H.Z. and C.U.I.W.; Data curation, X.C.; Formal analy-
sis, X.C., H.Z. and C.U.I.W.; Methodology, X.C., H.Z. and C.U.I.W.; Software, X.C.; Writing—original
Agriculture 2025, 15, 1627 24 of 27

draft, X.C., H.Z. and C.U.I.W.; Writing—review and editing, X.C., H.Z. and C.U.I.W. All authors have
read and agreed to the published version of the manuscript.

Funding: This research received no external funding.

Institutional Review Board Statement: Not applicable.

Data Availability Statement: The original contributions presented in this study are included in the
article. Further inquiries can be directed to the corresponding author(s).

Acknowledgments: We all acknowledge the support of Macao Polytechnic University (RP/FCHS-

02/2025).

Conflicts of Interest: The authors declare no conflicts of interest.

References
1. Wen, Z.; Chen, Y.; Liu, Z.; Meng, J. Biochar and Arbuscular Mycorrhizal Fungi Stimulate Rice Root Growth Strategy and Soil
Nutrient Availability. Eur. J. Soil. Biol. 2022, 113, 103448. [CrossRef]
2. Adomako, M.O.; Roiloa, S.; Yu, F.-H. Potential Roles of Soil Microorganisms in Regulating the Effect of Soil Nutrient Heterogeneity
on Plant Performance. Microorganisms 2022, 10, 2399. [CrossRef] [PubMed]
3. Chen, X.; Cui, F.; Wong, C.U.I.; Zhang, H.; Wang, F. An Investigation into the Response of the Soil Ecological Environment to
Tourist Disturbance in Baligou. PeerJ 2023, 11, e15780. [CrossRef]
4. Misbah, K.; Laamrani, A.; Khechba, K.; Dhiba, D.; Chehbouni, A. Multi-Sensors Remote Sensing Applications for Assessing,
Monitoring, and Mapping NPK Content in Soil and Crops in African Agricultural Land. Remote Sens. 2021, 14, 81. [CrossRef]
5. Chen, X.; Zhang, H.; Wong, C.U.I.; Li, F. Ecological Sensitivity Assessment of Central Plains Cities Using RS/GIS Technology: A
Case Study of the Zheng-Bian-Luo Urban Agglomeration. Pol. J. Environ. Stud. 2025, 34, 4633–4645. [CrossRef]
6. Chen, X.; Chen, F.; Cui, F.; Lei, W. Spatial Heterogeneity of Sustainable Land Use in the Guangdong–Hong Kong–Macao Greater
Bay Area in the Context of the Carbon Cycle: GIS-Based Big Data Analysis. Sustainability 2023, 15, 1715. [CrossRef]
7. Deng, L.; Mao, Z.; Li, X.; Hu, Z.; Duan, F.; Yan, Y. UAV-Based Multispectral Remote Sensing for Precision Agriculture: A
Comparison between Different Cameras. ISPRS J. Photogramm. Remote Sens. 2018, 146, 124–136. [CrossRef]
8. Zheng, H.; Cheng, T.; Li, D.; Yao, X.; Tian, Y.; Cao, W.; Zhu, Y. Combining Unmanned Aerial Vehicle (UAV)-Based Multispectral
Imagery and Ground-Based Hyperspectral Data for Plant Nitrogen Concentration Estimation in Rice. Front. Plant Sci. 2018, 9, 936.
[CrossRef] [PubMed]
9. Ge, X.; Ding, J.; Jin, X.; Wang, J.; Chen, X.; Li, X.; Liu, J.; Xie, B. Estimating Agricultural Soil Moisture Content through UAV-Based
Hyperspectral Images in the Arid Region. Remote Sens. 2021, 13, 1562. [CrossRef]
10. Sanogo, K.; Birhanu, B.Z.; Sanogo, S.; Ba, A. Landscape Pattern Analysis Using GIS and Remote Sensing to Diagnose Soil Erosion
and Nutrient Availability in Two Agroecological Zones of Southern Mali. Agric. Food Secur. 2023, 12, 4. [CrossRef]
11. Jain, S.; Sethia, D.; Tiwari, K.C. A Critical Systematic Review on Spectral-Based Soil Nutrient Prediction Using Machine Learning.
Environ. Monit. Assess. 2024, 196, 699. [CrossRef] [PubMed]
12. Escorcia-Gutierrez, J.; Gamarra, M.; Soto-Diaz, R.; Pérez, M.; Madera, N.; Mansour, R.F. Intelligent Agricultural Modelling of Soil
Nutrients and pH Classification Using Ensemble Deep Learning Techniques. Agriculture 2022, 12, 977. [CrossRef]
13. Wilhelm, R.C.; van Es, H.M.; Buckley, D.H. Predicting Measures of Soil Health Using the Microbiome and Supervised Machine
Learning. Soil. Biol. Biochem. 2022, 164, 108472. [CrossRef]
14. Ahmed, M.; Aslam, M.A.; Hayat, R.; Nasim, W.; Akmal, M.; Mubeen, M.; Hussain, S.; Ahmad, S. Nutrient Dynamics and
the Role of Modeling. In Building Climate Resilience in Agriculture: Theory, Practice and Future Perspective; Springer Nature:
Berlin/Heidelberg, Germany, 2022; pp. 297–316.
15. Kar, S.K.; Kumar, S.; Sankar, M.; Patra, S.; Singh, R.; Shrimali, S.; Ojasvi, P. Process-Based Modelling of Soil Erosion: Scope and
Limitation in the Indian Context. Curr. Sci. 2022, 122, 533–541. [CrossRef]
16. Villalobos, F.J.; Delgado, A.; López-Bernal, Á.; Quemada, M. FertiliCalc: A Decision Support System for Fertilizer Management.
Int. J. Plant Prod. 2020, 14, 299–308. [CrossRef]
17. Luo, L.; Li, Y.; Guo, F.; Huang, Z.; Wang, S.; Zhang, Q.; Zhang, Z.; Yao, Y. Research on Robust Inversion Model of Soil Moisture
Content Based on GF-1 Satellite Remote Sensing. Comput. Electron. Agric. 2023, 213, 108272. [CrossRef]
18. Singh, A.; Gaurav, K. Deep Learning and Data Fusion to Estimate Surface Soil Moisture from Multi-Sensor Satellite Images. Sci.
Rep. 2023, 13, 2251. [CrossRef]
19. Castaldi, F.; Hueni, A.; Chabrillat, S.; Ward, K.; Buttafuoco, G.; Bomans, B.; Vreys, K.; Brell, M.; van Wesemael, B. Evaluating the
Capability of the Sentinel 2 Data for Soil Organic Carbon Prediction in Croplands. ISPRS J. Photogramm. Remote Sens. 2019, 147,
267–282. [CrossRef]
Agriculture 2025, 15, 1627 25 of 27

20. Yuan, Y.; Gao, S.; Zhang, Z.; Wang, W.; Xu, Z.; Liu, Z. Edge-Cloud Collaborative UAV Object Detection: Edge-Embedded
Lightweight Algorithm Design and Task Offloading Using Fuzzy Neural Network. IEEE Trans. Cloud Comput. 2024, 12, 306–318.
[CrossRef]
21. Jiang, J.; Li, F.; Yang, J.; Kang, Z.; Li, J. Construction of Indoor Obstacle Element Map Based on Scene-Aware Priori Obstacle Rules.
ISPRS J. Photogramm. Remote Sens. 2023, 195, 43–64. [CrossRef]
22. Kurihara, J.; Nagata, T.; Tomiyama, H. Rice Yield Prediction in Different Growth Environments Using Unmanned Aerial
Vehicle-Based Hyperspectral Imaging. Remote Sens. 2023, 15, 2004. [CrossRef]
23. Yao, X.; Yi, Q.; Wang, F.; Xu, T.; Zheng, J.; Shi, Z. Estimating Rice Flower Intensity Using Flower Spectral Information from
Unmanned Aerial Vehicle (UAV) Hyperspectral Images. Int. J. Appl. Earth Obs. Geoinf. 2023, 122, 103415. [CrossRef]
24. Liu, K.; Wang, Y.; Peng, Z.; Xu, X.; Liu, J.; Song, Y.; Di, H.; Hua, D. Monitoring Soil Nutrients Using Machine Learning Based on
UAV Hyperspectral Remote Sensing. Int. J. Remote Sens. 2024, 45, 4897–4921. [CrossRef]
25. Piikki, K.; Wetterlind, J.; Söderström, M.; Stenberg, B. Three-Dimensional Digital Soil Mapping of Agricultural Fields by
Integration of Multiple Proximal Sensor Data Obtained from Different Sensing Methods. Precis. Agric. 2015, 16, 29–45. [CrossRef]
26. Söderström, M.; Sohlenius, G.; Rodhe, L.; Piikki, K. Adaptation of Regional Digital Soil Mapping for Precision Agriculture. Precis.
Agric. 2016, 17, 588–607. [CrossRef]
27. Mouazen, A.M.; Shi, Z. Estimation and Mapping of Soil Properties Based on Multi-Source Data Fusion. Remote Sens. 2021, 13, 978.
[CrossRef]
28. Islam, N.; Rashid, M.M.; Pasandideh, F.; Ray, B.; Moore, S.; Kadel, R. A Review of Applications and Communication Technologies
for Internet of Things (Iot) and Unmanned Aerial Vehicle (Uav) Based Sustainable Smart Farming. Sustainability 2021, 13, 1821.
[CrossRef]
29. Almalki, F.A.; Soufiene, B.O.; Alsamhi, S.H.; Sakli, H. A Low-Cost Platform for Environmental Smart Farming Monitoring System
Based on IoT and UAVs. Sustainability 2021, 13, 5908. [CrossRef]
30. Fei, S.; Hassan, M.A.; Xiao, Y.; Su, X.; Chen, Z.; Cheng, Q.; Duan, F.; Chen, R.; Ma, Y. UAV-Based Multi-Sensor Data Fusion and
Machine Learning Algorithm for Yield Prediction in Wheat. Precis. Agric. 2023, 24, 187–212. [CrossRef]
31. Nguyen, C.; Sagan, V.; Bhadra, S.; Moose, S. UAV Multisensory Data Fusion and Multi-Task Deep Learning for High-Throughput
Maize Phenotyping. Sensors 2023, 23, 1827. [CrossRef] [PubMed]
32. Ahmed, U.; Lin, J.C.-W.; Srivastava, G.; Djenouri, Y. A Nutrient Recommendation System for Soil Fertilization Based on
Evolutionary Computation. Comput. Electron. Agric. 2021, 189, 106407. [CrossRef]
33. Haddon, A.; Rapaport, A.; Roux, S.; Harmand, J. Model Based Optimization of Fertilization with Treated Wastewater Reuse. Adv.
Water Resour. 2023, 181, 104561. [CrossRef]
34. Sha, S.; Zhao, X.; Li, Y.; Li, C.; Zhu, L.; Wang, Y.; Gao, Q. Nutrient Expert System Optimizes Fertilizer Management to Improve
Potato Productivity and Tuber Quality. J. Sci. Food Agric. 2022, 102, 1233–1244. [CrossRef]
35. Tan, X.J.; Cheor, W.L.; Yeo, K.S.; Leow, W.Z. Expert Systems in Oil Palm Precision Agriculture: A Decade Systematic Review. J.
King Saud Univ.-Comput. Inf. Sci. 2022, 34, 1569–1594. [CrossRef]
36. Gao, J.; Zeng, W.; Ren, Z.; Ao, C.; Lei, G.; Gaiser, T.; Srivastava, A.K. A Fertilization Decision Model for Maize, Rice, and Soybean
Based on Machine Learning and Swarm Intelligent Search Algorithms. Agronomy 2023, 13, 1400. [CrossRef]
37. Li, S.; Liu, J.; Yang, Y.; Shen, R.; Jiang, J. Thinking as Human: Self-Reflective Reinforcement Learning Framework for Fertilization
Decision-Making. Smart Agric. Technol. 2025, 10, 100841. [CrossRef]
38. Das, P.P.; Singh, K.R.; Nagpure, G.; Mansoori, A.; Singh, R.P.; Ghazi, I.A.; Kumar, A.; Singh, J. Plant-Soil-Microbes: A Tripartite
Interaction for Nutrient Acquisition and Better Plant Growth for Sustainable Agricultural Practices. Environ. Res. 2022, 214, 113821.
[CrossRef]
39. Fung, E.; Wang, J.; Zhao, X.; Farzamian, M.; Allred, B.; Clevenger, W.B.; Levison, P.; Triantafilis, J. Mapping Cation Exchange
Capacity and Exchangeable Potassium Using Proximal Soil Sensing Data at the Multiple-Field Scale. Soil. Tillage Res. 2023,
232, 105735. [CrossRef]
40. Chen, X.; Zhang, H.; Wong, C.U.I. Spatial Distribution Characteristics and Pollution Evaluation of Soil Heavy Metals in
Wulongdong National Forest Park. Sci. Rep. 2024, 14, 8880. [CrossRef]
41. Huang, S.; Tang, L.; Hupy, J.P.; Wang, Y.; Shao, G. A Commentary Review on the Use of Normalized Difference Vegetation Index
(NDVI) in the Era of Popular Remote Sensing. J. For. Res. 2021, 32, 1–6. [CrossRef]
42. Xue, J.; Su, B. Significant Remote Sensing Vegetation Indices: A Review of Developments and Applications. J. Sens. 2017,
2017, 1353691. [CrossRef]
43. Jiang, Z.; Huete, A.R.; Chen, J.; Chen, Y.; Li, J.; Yan, G.; Zhang, X. Analysis of NDVI and Scaled Difference Vegetation Index
Retrievals of Vegetation Fraction. Remote Sens. Environ. 2006, 101, 366–378. [CrossRef]
44. Wang, F.; Gao, J.; Zha, Y. Hyperspectral Sensing of Heavy Metals in Soil and Vegetation: Feasibility and Challenges. ISPRS J.
Photogramm. Remote Sens. 2018, 136, 73–84. [CrossRef]
Agriculture 2025, 15, 1627 26 of 27

45. Wang, S.; Guan, K.; Zhang, C.; Lee, D.; Margenot, A.J.; Ge, Y.; Peng, J.; Zhou, W.; Zhou, Q.; Huang, Y. Using Soil Library
Hyperspectral Reflectance and Machine Learning to Predict Soil Organic Carbon: Assessing Potential of Airborne and Spaceborne
Optical Soil Sensing. Remote Sens. Environ. 2022, 271, 112914. [CrossRef]
46. Aasen, H.; Honkavaara, E.; Lucieer, A.; Zarco-Tejada, P.J. Quantitative Remote Sensing at Ultra-High Resolution with UAV
Spectroscopy: A Review of Sensor Technology, Measurement Procedures, and Data Correction Workflows. Remote Sens. 2018,
10, 1091. [CrossRef]
47. Ameer, S.; Cheema, M.J.M.; Khan, M.A.; Amjad, M.; Noor, M.; Wei, L. Delineation of Nutrient Management Zones for Precise
Fertilizer Management in Wheat Crop Using Geo-Statistical Techniques. Soil. Use Manag. 2022, 38, 1430–1445. [CrossRef]
48. Rezaee, L.; Davatgar, N.; Moosavi, A.A.; Sepaskhah, A.R. Implications of Spatial Variability of Soil Physical Attributes in
Delineating Site-Specific Irrigation Management Zones for Rice Crop. J. Soil Sci. Plant Nutr. 2023, 23, 6596–6611. [CrossRef]
49. Mathenge, M.; Sonneveld, B.G.; Broerse, J.E. Application of GIS in Agriculture in Promoting Evidence-Informed Decision Making
for Improving Agriculture Sustainability: A Systematic Review. Sustainability 2022, 14, 9974. [CrossRef]
50. Nath, S. A Vision of Precision Agriculture: Balance between Agricultural Sustainability and Environmental Stewardship. Agron. J.
2024, 116, 1126–1143. [CrossRef]
51. Peng, Y.; Zhong, W.; Peng, Z.; Tu, Y.; Xu, Y.; Li, Z.; Liang, J.; Huang, J.; Liu, X.; Fu, Y. Enhanced Estimation of Rice Leaf Nitrogen
Content via the Integration of Hybrid Preferred Features and Deep Learning Methodologies. Agronomy 2024, 14, 1248. [CrossRef]
52. Lambertini, A.; Mandanici, E.; Tini, M.A.; Vittuari, L. Technical Challenges for Multi-Temporal and Multi-Sensor Image Processing
Surveyed by UAV for Mapping and Monitoring in Precision Agriculture. Remote Sens. 2022, 14, 4954. [CrossRef]
53. Sharma, S.; Chatterjee, A. Comparing Soil Test Kits with Standard Lab-Based Soil Tests for Agricultural Soils. Crops Soils 2019, 52,
18–20. [CrossRef]
54. Uribeetxebarria, A.; Castellón, A.; Aizpurua, A. A First Approach to Determine If It Is Possible to Delineate In-Season n
Fertilization Maps for Wheat Using NDVI Derived from Sentinel-2. Remote Sens. 2022, 14, 2872. [CrossRef]
55. de Jesús Marcial-Pablo, M.; Gonzalez-Sanchez, A.; Jimenez-Jimenez, S.I.; Ontiveros-Capurata, R.E.; Ojeda-Bustamante, W.
Estimation of Vegetation Fraction Using RGB and Multispectral Images from UAV. Int. J. Remote Sens. 2019, 40, 420–438.
[CrossRef]
56. Khanal, S.; Kc, K.; Fulton, J.P.; Shearer, S.; Ozkan, E. Remote Sensing in Agriculture—Accomplishments, Limitations, and
Opportunities. Remote Sens. 2020, 12, 3783. [CrossRef]
57. Pande, C.B.; Moharir, K.N. Application of Hyperspectral Remote Sensing Role in Precision Farming and Sustainable Agriculture
under Climate Change: A Review. In Climate Change Impacts on Natural Resources, Ecosystems and Agricultural Systems; Springer
Nature: Berlin/Heidelberg, Germany, 2023; pp. 503–520.
58. Yu, H.; Kong, B.; Hou, Y.; Xu, X.; Chen, T.; Liu, X. A Critical Review on Applications of Hyperspectral Remote Sensing in Crop
Monitoring. Exp. Agric. 2022, 58, e26. [CrossRef]
59. Zhang, X.; Xiang, D.-Q.; Yang, C.; Wu, W.; Liu, H.-B. The Spatial Variability of Temporal Changes in Soil pH Affected by
Topography and Fertilization. Catena 2022, 218, 106586. [CrossRef]
60. Júnior, M.R.B.; Santos, R.G.D.; de Azevedo Sales, L.; de Oliveira, L.P. Advancements in Agricultural Ground Robots for Specialty
Crops: An Overview of Innovations, Challenges, and Prospects. Plants 2024, 13, 3372. [CrossRef]
61. Jafarbiglu, H.; Pourreza, A. A Comprehensive Review of Remote Sensing Platforms, Sensors, and Applications in Nut Crops.
Comput. Electron. Agric. 2022, 197, 106844. [CrossRef]
62. Bahlo, C.; Dahlhaus, P.; Thompson, H.; Trotter, M. The Role of Interoperable Data Standards in Precision Livestock Farming in
Extensive Livestock Systems: A Review. Comput. Electron. Agric. 2019, 156, 459–466. [CrossRef]
63. Salembier, C.; Segrestin, B.; Weil, B.; Jeuffroy, M.-H.; Cadoux, S.; Cros, C.; Favrelière, E.; Fontaine, L.; Gimaret, M.; Noilhan, C.;
et al. A Theoretical Framework for Tracking Farmers’ Innovations to Support Farming System Design. Agron. Sustain. Dev. 2021,
41, 61. [CrossRef]
64. Goldstein, A.; Fink, L.; Ravid, G. A Cloud-Based Framework for Agricultural Data Integration: A Top-down-Bottom-up Approach.
IEEE Access 2022, 10, 88527–88537. [CrossRef]
65. Munir, K.; Ghafoor, M.; Khafagy, M.; Ihshaish, H. AgroSupportAnalytics: A Cloud-Based Complaints Management and Decision
Support System for Sustainable Farming in Egypt. Egypt. Inform. J. 2022, 23, 73–82. [CrossRef]
66. Ju, C.; Son, H.I. Multiple UAV Systems for Agricultural Applications: Control, Implementation, and Evaluation. Electronics 2018,
7, 162. [CrossRef]
67. Merz, M.; Pedro, D.; Skliros, V.; Bergenhem, C.; Himanka, M.; Houge, T.; Matos-Carvalho, J.P.; Lundkvist, H.; Cürüklü, B.;
Hamrén, R.; et al. Autonomous UAS-Based Agriculture Applications: General Overview and Relevant European Case Studies.
Drones 2022, 6, 128. [CrossRef]
68. He, Q.; Zhao, H.; Feng, Y.; Wang, Z.; Ning, Z.; Luo, T. Edge Computing-Oriented Smart Agricultural Supply Chain Mechanism
with Auction and Fuzzy Neural Networks. J. Cloud Comput. 2024, 13, 66. [CrossRef]
Agriculture 2025, 15, 1627 27 of 27

69. Chen, B.; Su, Q.; Li, Y.; Chen, R.; Yang, W.; Huang, C. Field Rice Growth Monitoring and Fertilization Management Based on UAV
Spectral and Deep Image Feature Fusion. Agronomy 2025, 15, 886. [CrossRef]
70. Angelopoulou, T.; Balafoutis, A.; Zalidis, G.; Bochtis, D. From Laboratory to Proximal Sensing Spectroscopy for Soil Organic
Carbon Estimation—A Review. Sustainability 2020, 12, 443. [CrossRef]
71. Angelopoulou, T.; Tziolas, N.; Balafoutis, A.; Zalidis, G.; Bochtis, D. Remote Sensing Techniques for Soil Organic Carbon
Estimation: A Review. Remote Sens. 2019, 11, 676. [CrossRef]
72. Tziachris, P.; Aschonitis, V.; Chatzistathis, T.; Papadopoulou, M. Assessment of Spatial Hybrid Methods for Predicting Soil
Organic Matter Using DEM Derivatives and Soil Parameters. Catena 2019, 174, 206–216. [CrossRef]
73. Tripathi, A.; Tiwari, R.K.; Tiwari, S.P. A Deep Learning Multi-Layer Perceptron and Remote Sensing Approach for Soil Health
Based Crop Yield Estimation. Int. J. Appl. Earth Obs. Geoinf. 2022, 113, 102959. [CrossRef]
74. Leong, W.L.; Martinel, N.; Huang, S.; Micheloni, C.; Foresti, G.L.; Teo, R.S.H. An Intelligent Auto-Organizing Aerial Robotic
Sensor Network System for Urban Surveillance. J. Intell. Robot. Syst. 2021, 102, 33. [CrossRef]
75. Kalyani, Y.; Collier, R. A Systematic Survey on the Role of Cloud, Fog, and Edge Computing Combination in Smart Agriculture.
Sensors 2021, 21, 5922. [CrossRef] [PubMed]

Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to
people or property resulting from any ideas, methods, instructions or products referred to in the content.