Multimedia Tools and Applications (2024) 83:29505–29529

[Link]

A smart waste classification model using hybrid CNN‑LSTM

with transfer learning for sustainable environment

Umesh Kumar Lilhore1 · Sarita Simaiya2 · Surjeet Dalal3 · Robertas Damaševičius4

Received: 24 April 2023 / Revised: 24 August 2023 / Accepted: 27 August 2023 /

Published online: 13 September 2023
© The Author(s), under exclusive licence to Springer Science+Business Media, LLC, part of Springer Nature 2023

Abstract
Waste collection, classification, and planning have become crucial as industrialization and
smart city advancement activities have increased. A recycling process of waste relies on
the ability to retrieve the characteristics as it was in their natural position, and it reduces
pollution and helps in a sustainable environment. Recently, deep learning (DL) methods
have been employed intelligently to support the administration’s strategized waste manage-
ment and related procedure, including capture, classification, composting, and dumping.
The selection of the optimum DL technique for categorizing and forecasting waste is a
long and arduous process. This research presents a smart waste classification using Hybrid
CNN-LSTM with transfer learning for sustainable development. The waste can be classi-
fied into recyclable and organic categories. To classify waste statistics, implement a hybrid
model combining Convolutional neural networks (CNN) and long short-term memory
(LSTM). The proposed model also uses the transfer learning (TL) method, which incorpo-
rates the advantage of ImageNet, to classify and forecast the waste category. The proposed
model also utilises an improved data augmentation process for overfitting and data sam-
pling issues. An experimental analysis was conducted on the TrashNet dataset sample, with
27027 images separated into two classes of organic waste 17005 and recyclable waste 10
025 used to evaluate the performance of the proposed model. The proposed hybrid model
and various existing CNN models (i.e., VGG-16, ResNet-34, ResNet-50, and AlexNet)
were implemented using Python and tested based on performance measuring parameters,
i.e., precision, recall, testing and training loss, and accuracy. Each model was created with
a range of epochs and an adaptive moment estimator (AME) optimisation algorithm. For
the proposed method, the AME optimisation achieved the best optimisation and accuracy
and the least modelling loss for training, validation, and testing. The proposed model per-
formed the highest precision of 95.45%, far better than the existing deep learning method.

Keywords Smart waste · Classification · Sustainable development · Deep learning · CNN-

LSTM · Transfer learning

Extended author information available on the last page of the article

13
Vol.:(0123456789)
29506 Multimedia Tools and Applications (2024) 83:29505–29529

1 Introduction

Waste management applies to all such practices and activities needed to conquer from its
origins to closure. Waste can be in the form of a solid, liquid, or gas. Recycling refers to the
act of turning waste material back into new products. This idea frequently takes the recy-
cling of energy from discarded items into consideration. A recycling process of content
relies on retrieving the characteristics as it was in their natural position. It’s an environmen-
tally friendly substitute to “traditional” waste management, which can retain components
and reduce emissions of greenhouse gases. Reprocessing helps to decrease the consump-
tion of new resources and eliminate the loss of feasibly valuable resources, which reduces
energy consumption, environmental pollution, and air, water and land pollution.
Various procedures are employed to determine how to deal with all forms of waste,
including natural, manufacturing, and domestic. Household waste includes paperboard,
glass, plastic, sheets, and organic waste [1]. The absorption of non-recyclable landfill waste
all over the universe, along with the long and complex duration it takes for the majority
of its components to biodegrade, have a massive effect on our surroundings in the twenty-
first century if we, as little more than a society, do not act quickly to preclude this from
transpiring. Furthermore, among all of the best-known potential consequences to individu-
als, waste buildup can enable disease transmission by viruses like mosquitoes, flies, and
various other pests [2]. Regarding destroying the elegance of natural ecosystems, forest
destruction, and terrain profession to give sufficient area to open dumps, water and soil can
often be highly strung and heavily polluted because of the harmful pesticides in untreated
wastewater components. In addition, pollution can affect the local ecosystem, food supply,
farming, and industries, inevitably increasing illnesses and other problems for people and
the planet’s ecological processes [3].
Sustainable disposal of waste attempts to reduce the volume of solid garbage dumped
in landfills or by burn. The hierarchical structure of waste management focuses on pre-
vention, reduction, recycling, composting, recovering energy, and, subsequently, cure
for waste disposal, which is the foundation of a sustainable waste management strategy.
Waste accrual has become a more severe problem in the past few years for three principal
reasons. The first concern is the limitation of recyclable waste items, although industries
have been working to develop more environmentally friendly and renewable materials
for a long time [4].
The second main factor is overpopulation, which is currently one of the most prominent
issues. The reality that a substantial number of individuals require a wide range of materi-
als implies a very complicated logistical problem when attempting to handle waste genera-
tion, resulting in a higher percentage of merchandise that can be reprocessed but instead
ends up in an ocean and landfill, impacting the families and communities of millions of
marine animals. At last, the third possibility is our lack of engagement as a civilization in
climate change or other environmental issues [5].
Waste management is a critical global phenomenon. Since such an earth’s population
and requirement of higher living standards criteria increase, so does the level of waste pro-
duced [7]. Researchers are becoming greatly worried about waste processing and its possi-
ble repercussions and therefore are attempting to find solutions. The method of transform-
ing industrial waste into new substances and entities is called "Recycling" [6]. To address
these waste management challenges, some innovative, cutting-edge techniques such as
Artificial intelligence (AI), Machine Learning (ML), and Deep Learning (DL). Among the
most famously accepted methods in environmental studies is the DL technique. DL makes

13
Multimedia Tools and Applications (2024) 83:29505–29529 29507

a model capable of learning complicated selective activities to help agro-industrial waste

management [7]. A DL method is successfully utilised in the Sustainable waste manage-
ment research area. DL is a broad subset of ML techniques, including various analytical
tools and strategies for implementing ANN with object classification. Consequently, differ-
ent CNN approaches have been incorporated into studies to detect and locate unauthorized
dumps utilising roadside object data and extremely high-resolution aerial image graphs [7].
Many renowned researchers worldwide have made significant efforts to achieve Sustain-
able waste management utilising DL methods. SWM has widely used DL features to solve
various problems, including waste identifiers and separation, actual bin-level sensing, and
waste management forecasting. DL approaches are capable of identifying and learning fea-
tures actively from images. This distinguishing feature has significantly improved image
recognition and tracking [8].
A deep learning method, i.e., CNN, performs excellently in image identification and cat-
egorization assignment, i.e., Waste classification. However, the most significant problem
with the CNN model occurs within model training because of the gradient disappearing
issues and sometimes gradient explosions issues along with data loss, which may cause
the model’s poor performance. To overcome this issue in this research, we utilize a CNN
model with LSTM and Transfer learning. Both temporal and spatial dimensions can be
used to teach characteristics to a CNN. An LSTM network analyses data from sequences
by cycling through time steps and discovering the long-term relationships among them.
A CNN-LSTM network employs convolution and LSTM combined layers to acquire
knowledge from the training information. That can be more efficient for image data pre-
diction and classification [9]. Currently, The classification of waste is fraught with several
difficulties.

• Complexity: It stems from the fact that waste materials are extremely diverse in their
composition, physical properties, and possible environmental effects. Accurately clas-
sifying those calls for both specialized knowledge and in-depth investigation.
• Subjectivity: It refers to the fact that waste classification can sometimes be open to
debate and differ from one regulatory body or expert to another.
• Insufficient Data: Inaccurate waste classification can be hampered by insufficient data
about the waste materials.
• Traditional methods: of waste classification can be time-consuming and expensive,
particularly for large volumes of waste. This is especially true when the volume of
waste is large.
• Emerging Waste Streams: New types of waste, such as electronic waste and certain
plastics, present new challenges when classifying them properly because of the com-
plex nature of their components.

Deep learning techniques, which are a subfield of artificial intelligence, have the poten-
tial to improve waste classification and find solutions to problems. In this paper, we pro-
posed a hybrid model focused on CNN-LSTM, which further uses ImageNet weights
with a transfer learning method to categorize waste data. Furthermore, we cover the pro-
posed hybrid Model CNN-LSTM technique with TL and the datasets utilised to evaluate
the hybrid model in recognizing and classifying waste categories, such as recyclable and
organic waste. The proposed model’s performance is tested on the Kaggle waste data-
set and compared with existing CNN models (i.e., VGG-16, ResNet-34, ResNet-50, and
AlexNet). Each model was created with a range of epochs and an adaptive moment estima-
tor (AME) optimisation algorithm [10].

13
29508 Multimedia Tools and Applications (2024) 83:29505–29529

The complete research article is organized as follows: Section 2 covers the related work
in waste recognition and classification by various machine learning deep learning-based
models. Section 3 covers the materials and methods, Section 4 covers the experimental
results and discussion and Section 5 covers the conclusion and future work.

2 Related work

Waste management plays an important role in the sustainable development of the country.
Waste management is a hot area of research. This section covers the analysis of existing
research in waste management.
A deep learning-based hybrid waste classification model is presented in [11]. They
divided garbage into seven different groups employing SVM with CNN deep learning
model. The proposed hybrid SVM-CNN model performed better than the existing CNN
with 83% precision. RecycleNet, a framework that can precisely optimize waste, is pre-
sented in [12]. This model uses a sophisticated CNN architecture for certain recyclable
product categories. The proposed system altered the existing linked patterns and skipped
the dense block patterns. Although the proposed RecycleNet received a score of 87.58% in
precision in the TrashNet image data. The proposed model also decreased the number of
characteristics from seven million to approximately three million.
Applying the ResNet architecture [13] introduced a DNN-TC approach that intelligently
categorizes waste in smart waste sorting equipment. To reduce complexity, two completely
interconnected layers were added to the ResNeXt-101 general framework. Further, the pre-
trained weights of ResNeXt-101 from the ImageNet database [14] were imported through-
out the training phase, resulting in a better outcome of 94% precision over the TrashNet
data. The DNN-TC method achieves better than ResNeXt-101 for identifying plastic and
recycling in their test. However, the author does not explain why DNN-TC performs worse
for metal, paper, and plastic when compared to ResNeXt-101.M-b Xception, a unique con-
nection expansion-based network optimisation approach for garbage digital image recogni-
tion, which was introduced in [15]. The TrashNet collection produced the best outcomes
94.34% accuracy rate. While their system boosts resilience and accuracy rate compared to
earlier approaches, it also raises processing costs and variable potential.
A deep CNN categorizes the local features from the waste image file [16]. To increase
classification performance, the characteristics of a pre-trained classifier are altered in this
article to use the waste dataset after it has been introduced and pre-trained upon this Ima-
geNet massive image dataset. To classify waste, several investigators have developed modi-
fied CNN models. Hybrid CNN architecture was designed in [17] and assessed the model’s
effectiveness using two open waste datasets. A wide range of typical CNN techniques was
evaluated. A VGGNet-19 achieved the highest precision at 92.9%. The proposed CNN
technique performed faster and only managed a 90% accuracy rate.
Linen, plastics, biological refuse, and glassware waste were the four class descriptors
used by [18] to assess an unnamed Deep CNN with 4600 self-acquired waste photos. They
obtained an f-score of 69% and 85% for each category. A modified CNN-based waste cat-
egorization framework is offered in [19] to categorize organic and inorganic waste. The
gathered waste can be divided into various categories using a flap. After that, the comput-
erized order aids in the time employed on cleanliness. The organizations can simultane-
ously organize and manage the generated waste too. An image enhancement and a modi-
fied CNN for waste classification frameworks were addressed in [20]. Researchers executed

13
Multimedia Tools and Applications (2024) 83:29505–29529 29509

research primarily focused on finding organic waste. The proposed model achieved 87.98%
accuracy, which was the best compared to existing CNN models.
The expanding world accomplishment of reducing emissions has made recycling waste
identification and classification a crucial step in advancing the growth of a sustainable
society. A deep learning-based Model is discussed in [16] for a smart classification of
recyclable material. The proposed model utilises ResNet and Transfer learning. The pro-
posed model achieved an accuracy of 90.17% over the traditional CNN model. The method
for determining if an image includes waste or doesn’t was created by [17]. They made a
trash app for Android mobile devices. A CNN programme recognizes the trash area of the
image termed GarbNet. After GarbNet optimisation, precision was 87.69%, and specificity
was 93.45%. The research also accessed the GarbageIn images database, a waste-sensitive
repository comprising real-world, geotagged photos. Table 1 presents a summarized analy-
sis of recent research in the field of waste management.
Compared to the findings of previous studies on this domain shown in Table 1, these
models cannot gain the maximum level of accuracy in predicting waste. In the work cited
by [11], the various machine learning models, i.e., DenseNet-169 and XGB, are being
applied with 88.9%. Meanwhile, [20] tried RNN with Agriculture Waste with a 91.3%
accuracy level. ResNet50 Model was integrated into [13] on Household Waste but gained
91.89% accuracy. In [39], the authors applied the ND-CNN (GoogleNet) to Industrial
Waste but did not achieve a sufficient accuracy level (92.48%).

3 Materials and methods

This section covers the discussion on existing methods and the proposed method.

3.1 Dataset

This research utilises an online Kaggle waste dataset [31]. The dataset consists of 27025
images separated into two classes of organic waste (O: 17005) and recyclable waste (R:
10020) was used to test the proposed model. An improved data augmentation technique
was applied to enhance the waste image quality. Considering recyclable waste in the envi-
ronment is random and irregular, measures of colour space transformation, rotation, flip-
ping, and noise injection have been taken to augment the image sample [32]. The details
are discussed in Section 3.4.1. Figure 1 presents the image categories and sample images
in the dataset.

3.2 Proposed hybrid model

This research presents a smart waste classification using Hybrid CNN-LSTM with trans-
fer learning for sustainable development. The waste can be classified into recyclable and
organic categories. Implement a hybrid model that combines CNN with LSTM to classify
waste statistics. Transfer learning is also used by the proposed model, which incorporates
the advantage of ImageNet, to classify and forecast the waste category [21]. Figure 2 pre-
sents the architecture of the proposed hybrid model.

13
 29510

13
Table 1  Summary of waste management research
Reference Model used Waste type Waste dataset Accuracy result Transfer learning Optimisation algorithm

[11] DenseNet-169 and XGB Industrial Waste Kaggle dataset 88.9% Im-Break Chimp optimisation
[12] RNN Agriculture Waste Self-made dataset 91.3% Im-Break Shuffled Frog-Leaping
[13] ResNet50 Model Household Waste Self-made dataset 91.89% Im-Break Stochastic gradient descent
[14] AlexNet Model and VGG16 Industrial Waste Kaggle dataset 89.98% Im-Break Genetic
Model
[15] GoogleNet Plastic and liquid waste Ocean Waste dataset 88.9% None Gradient Descent
[16] ResNet-152 Agriculture waste Kaggle waste dataset 87.78% None Adaptive Learning Rate
[17] Darknet NN Household and Industrial Waste Self-made dataset 89.9% None Genetic
[18] Lightweight NN Water waste Kaggle waste dataset 91.78% None Gradient Descent
[19] ResNet-152 Solid, Liquid Waste Self-made dataset 89.21% None Adaptive Learning Rate
[20] ND-CNN(GoogleNet) Industrial Waste Kaggle l Waste data 92.48% ImageNet Genetic
Proposed model CNN-LSTM with Transfer Solid waste Kaggle waste dataset 95.45% ImageNet AME
learning
Multimedia Tools and Applications (2024) 83:29505–29529
Multimedia Tools and Applications (2024) 83:29505–29529 29511

Fig. 1  Sample images in the dataset

Fig. 2  Architecture of proposed hybrid model

3.3 Working of proposed hybrid model

This proposed model primarily consists of two segments: a separate autonomous RNN
component and a CNN component with an image size of (224*224). In the proposed
model, an independent, autonomous RNN component has two different LSTM layers, both
of which are of batch shape 2048. However, the CNN is finally carried via a pre-trained
TL model, i.e. ResNet-50, InceptionResNet-V2, until it achieves the final destination layer,
i.e., Convolutional layer, that includes bottleneck functionalities; these all contain a batch
size of 64 [22]. Figure 3 presents the working of the proposed Hybrid CNN-LSTM with
the TL model.

13
29512 Multimedia Tools and Applications (2024) 83:29505–29529

Fig. 3  Working of proposed hybrid CNN-LSTM with transfer learning

The components’ results are combined through a variable size-based multiplication pro-
cess. The result is procured further into the activation function (Soft-Max) and the cat-
egorisation layer of different classes for multi-class classifications. In the next phase, all
the Convolutional layers with various filter strengths can recognise waste images with mag-
nifying factors because every convolution operation may take images of a specific length
[23]. The proposed hybrid model contains the following key components. a) LSTM layer
with dense feature; b) SoftMax activation function c) fully connected layer component; e)
Convolutional layer; f) pooling layer. The complete working of the model is as follows.

3.3.1 Improved data augmentation phase

To deal with data unbalancing and overfitting data augmentation phase is used. We utilise
an improved data augmentation process in two steps, i.e., background modification (Back-
Mod) and "ImageDataGenerator()" [24]. Figure 4 presents the architecture of the proposed
improved data augmentation model.
The first phase performs an image background modification (BackMod). The idea
besides this technique is to contextualise existing, classified waste. BackMod mainly analy-
ses noise and other issues throughout the waste sample. The images within context sub-
sequently allow waste analysers to concentrate on the aspects related to the waste images
[25]. Figure 5 presents a data augmentation background enhancement process for waste
images.
In the second phase, we undertake a multi-scale data augmentation process for the
training waste data sample. Using a random combination of key features with a set of
random methods, i.e., object rotation, flipping in both vertical and horizontal directions,
object translation and intensity variation, this phase prevents unbalanced class and overfit-
ting issues. A built-in Python Keras module, "ImageDataGenerator ()", was employed to
achieve the augmentation feature on the image dataset [26, 27]. Figure 6 presents Data
augmentation image flipping for recyclable (R) waste images.
Image frames were flipped in various angles, rotated vertically and horizontally, and
the centre feature was selected as "True." We employed a data over-sampling technique
combining the data augmentation by phases 1 and 2 to balance waste visuals of every
class because the categories of waste statistics are unbalanced and reflect a Gaussian
distribution [28].

13
Multimedia Tools and Applications (2024) 83:29505–29529 29513

Fig. 4  Architecture of improved data augmentation model

Fig. 5  Data augmentation background enhancement for waste images (Recyclable). a Input image, b pro-
cess image and c output image

3.3.2 Parameter optimisation

The proposed model utilises an adaptive moment estimator (AME) based optimisation. AME
becomes the next strategy that calculates adaptive learning scores of every factor. AME main-
tains an exponential decay mean of previous variations (V), comparable to motion, in contrast
to an exponential decay mean of squared gradients (S). AME operates most like a massive ball
experiencing internal friction instead of motion and can be visualised as a ball rolling down-
wards. As a result, AME favours a smooth minimum in the error function [29, 30].

(1)
( ) { )
Vt = (𝛾 1∗ Vt−1 + (1 − 𝛾 1 ) ∗ gt }

(2)
( ) { )
St = (𝛾 2∗ St−1 + (1 − 𝛾 2 ) ∗ gt2 }

13
29514 Multimedia Tools and Applications (2024) 83:29505–29529

Fig. 6  Data augmentation image flipping for recyclable (R) waste images

where V and S are the vectors, gt Gradient descent, 𝛾 optimisation. AME method helps to
fine-tune the hyperparameters and improve prediction accuracy. DL methods were trained,
optimised, and fine-tuned using hyperparameters. To avoid overfitting during the initial
training phase and fine-tuning the parameters, we utilised feature check-pointers, callbacks,
and loss-tracking features, and the optimal feature model’s weights got retained [31]. Opti-
misation techniques were used to finalise the proposed model, i.e., training rates, chang-
ing epochs, and hyperparameters with the degree of tolerance "1e-2". Next, to identify
the waste classes, we utilise a fine-tuned model created employing these ideal parameters.
Table 2 presents the parameters used to implement the proposed model [32].

Table 2  Parameters for CNN model (Pre-trained model)

Parameters Details Value

Window size Used for pooling layer 2

No. of epochs Used for epoch cycle 100 and 200
No layers in LSTM Used by LSTM model 2
Batch size To proceed with the data 32
Activation function Used for model activation SOFTMAX function
Learning rate It defines model learning 0.001
Batch normalisation Used to Normalisation of batches Yes
Optimizer Used for Optimisation Adaptive moment estimator

13
Multimedia Tools and Applications (2024) 83:29505–29529 29515

3.3.3 LSTM module

The LSTM scheme was created by [33] to address the problems associated with the train-
ing process, which contains lengthy dependencies that arise from the gradient constraints
in CNN. An LSTM model consists of several sequentially linked sub-networks called
"memory blocks". The block is designed to maintain its status over time despite controlling
the overall data stream via "non-linear gating" components. Figure 7 presents the working
of the LSTM model. An LSTM model contains various flags, i.e., yield value h (t), input X
(t). In LSTM architecture, all the output blocks are recurrently linked with all the gates and
input blocks.
An LSTM computing is achieved by mapping all the input sequence X = ­(X1….Xn) with
all the output Y = ­(Y1, ……Yn). The computing equations are as follows:

(3)
[ )
f (t) = γ[Wf ∗ (ht−1 , xt + bf ]

(4)
[ )
i(t) = γ[Wi ∗ (ht−1 , xt + bi ]

(5)
[ )
O(t) = γ[Wo ∗ (ht−1 , xt + bo ]

(6)
[ )
C(t) = f (t) ∗ C(t−1) + i(t) ∗ (tan(h)[WC ∗ (ht−1 , xt + (bC )]

h(t) = O(t) ∗ {(tan(h)[C(t)]} (7)

γO = γ[Wo ∗ (ht−1 , xt + (bo + bf + bi )] (8)

[ )

In the above equations (Eqs. 1 to 4) i (t) denotes input activation function, f (t) For-
get Gates and O (t) Output gates. Similar parameters b­ f, ­bi, and b­ o represents a bias value
towards a forget Gate’s input and output. Also ­Wf, ­Wi, and ­Wo represents weigh parameters

Fig. 7  LSTM model

13
29516 Multimedia Tools and Applications (2024) 83:29505–29529

towards Forget Gates input and output. Equation 5 represents the calculation of hidden
layer outputs. Also, Eq. 6, h (t) denotes a hidden layer, γ represents a sigmoid activation
function [34, 35].

3.4 Transfer learning

It is a deep learning approach that employs the ideas of field and assignment to transmit
knowledge amongst multiple physical conditions. It has a margin distribution function M ­ x
(x,y) among a source domain (­ Sd), Image labelling (­ IL), and a source component (­ Sc), Ima-
geNet ­(Is) [36, 37]. Additionally, there is an objective domain ­(Od), a Waste database ­(Wd),
and a target activity ­(Ta), an appropriate label, for the identification and classification of
all the possible recyclable waste ­(Rd). With the knowledge acquired from I­s and ­IL, the
­TL method aims to learn the objective probability distribution ­OPT(x,y) in ­Wd. An Ima-
geNet model contains over 12 million images with 1000 + classes in different categories. In
this research for the classification of recyclable waste, we applied TL models with hybrid
CNN-LSTM, and it helps to select pre-trained characteristics of the ImageNet and achieve
higher accuracy [38–40].

3.5 Performance measuring parameters

To measure the performance of the proposed model and various existing CNN model fol-
lowing performance-measuring parameters was calculated [41–43]].
[TruePositive]
Precision = (9)
[TruePositive + FalsePositive]

[TruePositive]
Recall = (10)
[Orignallypostivedata]

[Precision ∗ Recall]
F1 − Score = 2 ∗ (11)
[Precision + Recall]

[TruePositive + TrueNegative]
Accuracy = (12)
[TruePositive + TrueNegativeon + FalsePositive + FalseNegative]

Besides the above measures, the following parameters are widely used for performance
analysis.

• Training loss: A training loss evaluates the model’s performance across the training
dataset and can be utilised to determine how well a deep neural network algorithm
accurately reflects the training sets. A training loss is typically computed mathemati-
cally by summing the losses towards the training iteration [44].
• Validation loss: A similar measure, called a validation loss, is utilised to evaluate how
well a neural network-based model performed on validation data. A validation loss is
determined from the aggregate losses for each sample within validation data and is
comparable to the training loss [45].

13
Multimedia Tools and Applications (2024) 83:29505–29529 29517

4 Experimental results and discussion

This section covers the experimental details and results analysis of the proposed hybrid
model and existing CNN models, i.e., ResNet-50, AlexNet, ResNet-34 and VGG-16 mod-
els [40] at the waste dataset.

4.1 Experimental details

The implementation of the proposed Model and the existing Model was performed using
Python programming. The hardware and software features in the implementation and oper-
ation include "Nvidia-GeForce" Graphics, "Google-Collaborator" with 52 GB of primary
memory, Python, python libraries, and Linux 64-bit OS. For DL, the packages cuDNN-
7.6.5, tensor-flow-V2.11.0, CUDA-V10.0 and Keras-V2.10 were utilised. For visualization,
we mainly used matplotlib and Sea-born libraries [46–48].
Experimental results were performed for the proposed model and existing VGG-16,
ResNet-34, ResNet-50, and AlexNet, using the TrashNet waste dataset. The dataset con-
sists of 27025 images separated into two classes of organic waste (O: 17005) and recy-
clable waste (R: 10020) was used to test the proposed model. After fine-tuning the image
parameters, the dataset was divided into 3 classes (Training: Validation: and testing).
An experimental analysis was performed for two dataset divisions. Experiment one used
the dataset ratio of 80: 10: 10 (Training: Validation: Testing) and the number of epochs
from 0-50. Similar to experiment 2 dataset ratio of 70: 15: 15 (Training: Validation: Test-
ing) and the number of epochs from 0-100 were used. Table 3 presents an experimental
parameters overview for the proposed model, and Figure 8 shows the sample experimental
results (Organic and Recyclable).

Experiment 1 The dataset consists of 27025 images separated into two classes of organic
waste (O: 17005) and recyclable waste (R: 10020) was used to test the proposed model.
After fine-tuning the image parameters, the dataset was divided into 3 classes (Training:
Validation: and testing). In experiment one, the dataset ratio was 80: 10: 10 (Training: Vali-
dation: Testing), and the number of epochs was 0, 30, and 50. The experimental results for
experiment one are as follows.

Table 3  Experimental parameters for the proposed model

Model Layer (type) Output details Parameters

Conv-1d ( None, 28, 64) 192

MaxPooling-1d ( None, 14, 64) 0
CNN Flatten ( None, 896) 0
Dense_4 ( None, 50) 44,850
None_1 ( None, 1) 51
Total parameters of CNN 45093 Trainable: 45093 Non-Trainable: 0
LSTM_7 (None, 50) 16800
LSTM Droup_Out_5 (None, 50) 0
Dense_16 (None, 25) 1275
Dense_17 (None, 1) 26
Total parameters of LSTM 18101 Trainable: 18101 Non-Trainable: 0

13
29518 Multimedia Tools and Applications (2024) 83:29505–29529

Fig. 8  Experimental results (organic and recyclable)

Training and validation accuracy and loss were calculated for dataset 80: 10: 10
(training: validation: validation: testing). Figure 9a presents training and validation
accuracy results for the proposed model from 0–30-50 epochs. When the numbers of
epochs are less, the training and validation accuracies are also less, but once the number
of epochs increases, the accuracy results are increased for training and validation from
50 to 95%. A model with higher training and validation accuracy is always in demand,
and the proposed model is fit. Figure 9b presents training and validation loss results for
the proposed model from 0–50 epochs. Initially, the loss results are higher for epochs
0–30, but once we increase it up to 50, the loss % is also decreased, a fewer loss results
show better performance.
Figure 10 presents a ROC graph of waste classification for experiment 1 for 50
epochs. A graph called the ROC represents how well an algorithm for classification
works across all benchmarks. The graph is plotted among TPR and FPR. The ROC
graph was plotted for various waste types, i.e., cardboard, glass, metal, plastics, and
paper waste.

Fig. 9  a Training and Validation accuracy and b Training and Validation loss for the proposed model for
0–50 Epochs and dataset 80: 10: 10 (training: validation: testing)

13
Multimedia Tools and Applications (2024) 83:29505–29529 29519

Fig. 10  ROC graph of proposed model for waste classification for 50 epochs

Figure 11 presents a Precision-Confidence (TPR) Graph for all waste classes. The graph
is plotted among precision value and confidence results. A higher precision shows better
performance. The graph is plotted for various waste types, i.e., carboard-0, glass-1, metal-
2, plastics-3, and all classes-4.

Fig. 11  Precision-confidence
graph for all waste classes

13
29520 Multimedia Tools and Applications (2024) 83:29505–29529

Fig. 12  Experimental results

of the proposed model for the
F1-Measure graph

Figure 12 presents an F1-Measure graph for all waste classes. The graph is plot-
ted among the F1-score value and confidence results. A higher precision shows better
performance. The graph was plotted for various waste types, i.e., carboard-0, glass-1,
metal-2, plastics-3, and all classes-4.
Table 4 presents the error rate and total time comparison for the existing and pro-
posed Models for 0, 30, and 50 epochs. The proposed model takes 4589 s for 30 epochs,
and its error rate was also the lowest at 0.019578. Similar to the 50 epoch’s error rate
was 0.016578%, and the total time was 7586 s. A model with less error rate and time is
always in demand; the proposed model achieved both factors. The proposed model takes
4589 s for 30 epochs, and its error rate was also the lowest at 0.019578. Similar to the
50 epoch’s error rate was 0.016578%, and the total time was 7586 s. A model with less
error rate and time is always in demand; the proposed model achieved both factors.
Similarly, Table 5 presents an experimental result comparison of experiment 1 for
existing and proposed Models, i.e., Precision, Recall, F1-Score and Accuracy for 0–50
epochs and the [Link] 10 dataset. The proposed model achieves a precision of 95.4%,
recall of 94.3%, F1-score 93.1% and accuracy of 95.7%, which is best compared to all
the existing methods.

Table 4  Error rate and total time comparison for existing and proposed model
Model Transfer Division details 30 epochs Total time 50 epochs Total time
learning (training: validation: error rate (in seconds) error rate (in seconds)
testing) (%) (%)

VGG-16 ImageNet 80: 10: 10 0.078871 7898 0.05159 13193

ResNet-34 ImageNet 80: 10: 10 0.071847 5478 0.062492 10172
ResNet-50 ImageNet 80: 10: 10 0.052381 6898 0.05717 11093
AlexNet ImageNet 80: 10: 10 0.295718 4789 0.18788 8589
Proposed model ImageNet 80: 10: 10 0.019578 4589 0.016578 7568

13
Multimedia Tools and Applications (2024) 83:29505–29529 29521

Table 5  Precision, recall, Model Precision Recall F1-score Accuracy

F1-score, accuracy results
comparison for existing and
VGG-16 models 0.861 0.883 0.891 0.889
proposed model for 50 epochs
and [Link] 10 (training: ResNet-34 0.883 0.892 0.906 0.891
validation: testing) dataset ResNet-50 0.897 0.893 0.895 0.902
AlexNet 0.912 0.925 0.918 0.917
Proposed model 0.954 0.943 0.931 0.957

Experiment 2 The dataset consists of 27025 images separated into two classes of organic
waste (O: 17005) and recyclable waste (R: 10020) was used to test the proposed model.
After fine-tuning the image parameters, the dataset was divided into 3 classes (Training:
Validation: and testing). In experiment two, the dataset ratio was 70: 15: 15 (Training: Val-
idation: Testing), and the number of Epochs was 0–50-100. The experimental results for
experiment 2 are as follows.

For dataset 70: 15: 15 (training: validation: testing), and epoch 0–50-100, training, vali-
dation accuracy and loss were calculated in experiment 2. Figure 13a presents training and
validation accuracy results for a proposed model from 0–50 epochs. When the numbers of
epochs are less, the training and validation accuracies are also less, but once the number of
epochs increases, the accuracy results are increased for training and validation from 75%
to 95.8%. A model with higher training and validation accuracy is always in demand, and
the proposed model is the best fit. Figure 13b presents training and validation loss results
for the proposed model from 0–50 epochs. Initially, the loss results are higher for epochs
0–20, but once we increase it to 50, the loss % also decreases. Fewer loss results show bet-
ter performance.
Precision and recall results were calculated for 100 Epochs and dataset 70: 15: 15 (train-
ing: validation: testing). Figures 14 and 15 present the proposed Model for Precision and
Recall experimental results. Graph 14 was plotted for precision Vs confidence, and Graph
15 for recall and confidence. A model which generates higher precision and recall can be
treated as best.

Fig. 13  Experimental results of the proposed model for Epoch 0–50 and dataset 70: 15: 15 (training: valida-
tion: testing), a training and validation accuracy and b training and validation loss

13
29522 Multimedia Tools and Applications (2024) 83:29505–29529

Fig. 14  Experimental results of

the proposed model for precision
value (100 Epochs and dataset
70: 15: 15 (training: validation:
testing)

Fig. 15  Experimental results of

the proposed model for recall
(100 epochs and dataset 70: 15:
15 (training: validation: testing)

For dataset 70: 15: 15 (training: validation: testing), with 100 epochs, various perfor-
mance measuring parameters, i.e., precision, recall, F-measure, training-loss, and vali-
dation-loss, were calculated for the proposed model. Figure 16 presents an experimental
graph for all these results.
Table 6 presents the error rate and total time comparison for the existing and pro-
posed Models for 50 and 100 epochs. The proposed model takes 4508 s for 50 epochs,
and its error rate was also the lowest at 0.018181. Similar to the 100 epoch, the error
rate was 0.032578, and the total time was 8508 s. A model with less error rate and time
is always in demand; the proposed model achieved both factors. The choice of optimizer
can significantly impact the accuracy of the proposed model. Different optimizers have
strengths and weaknesses, and their performance can vary depending on the dataset’s
characteristics and the specific use case. Here are some general observations about the
impact of different optimizers on the accuracy of the proposed:

13
Multimedia Tools and Applications (2024) 83:29505–29529 29523

Fig. 16  Experimental results of the proposed model for various training and validation loss (for Epoch
0–100)

Table 6  Error rate and total time comparison for existing and proposed model
Model Transfer learn- Division details 50 epochs Total time 100 epochs Total
ing (training: vali- error rate (in seconds) error rate time (in
dation: testing) (%) (%) seconds)

VGG-16 ImageNet 70: 15: 15 0.04101 12103 0.08191 25193

ResNet-34 ImageNet 70: 15: 15 0.05287 10122 0.102107 20172
ResNet-50 ImageNet 70: 15: 15 0.4318 11901 0.851018 21091
AlexNet ImageNet 70: 15: 15 0.1334 8280 0.25414 15580
Proposed model ImageNet 70: 15: 15 0.018181 4508 0.032578 8508

• SGD is a simple optimizer that can work well for small datasets or simple models.
However, it can slowly converge and get stuck in local minima.
• Adam is a popular optimizer that works well for most deep learning problems, includ-
ing proposed. It can converge faster than SGD and can handle noisy or sparse gradients.
However, it can be sensitive to the choice of hyperparameters and can sometimes over-
fit the data.
• Adagrad can work well for sparse data and can converge quickly. However, it can accu-
mulate too much historical information and cause the learning rate to decay quickly,
leading to slow convergence.
• Adadelta is similar to Adagrad, but it can handle the decay problem using a moving
window of the past gradients. It can converge faster than Adagrad and requires fewer
hyperparameters.
• RMSProp is a good optimizer for deep learning models, including the proposed model.
It can handle the problem of vanishing or exploding gradients and adapt to different
learning rates for different parameters. However, it can converge slowly and require
careful tuning of the hyperparameters.

13
29524 Multimedia Tools and Applications (2024) 83:29505–29529

Table 7 presents an experimental result comparison for existing and proposed Models,
i.e., Precision, Recall, F1-Score and Accuracy. The proposed model achieves a precision of
95.4%, recall of 94.3%, F1-score 93.1% and accuracy of 95.7%, which is best compared to
all the existing methods. Figures 9, 10, 11, 12, 13, 14, 15 and 16 presents the experimental
results for the proposed hybrid and existing models.

4.2 Discussion

This research used a deep Convolutional neural network method to categorise waste into
multiple classes. This research used a hybrid model that included CNN-LSTM and a
trained TL model. Experimental results for proposed and existing models are as follows.
The results were calculated in two experiments. Experiment 1 with 100 epochs and 70: 15:
15 (training: validation: testing) dataset. Table 3 presents the experimental parameters for
the proposed model.
In experiment 1, the dataset consisting of 27025 images separated into two classes of
organic waste (O: 17005) and recyclable waste (R: 10020) was used to test the proposed
model. After fine-tuning the image parameters, the dataset was divided into 3 classes
(Training: Validation: and testing). On the same dataset, the splitting ratio was 80: 10: 10
(Training: Validation: Testing) and the number of Epochs are 0–30 and 50.
Figure 9 presents Training and Validation Accuracy and Training and Validation loss
for the proposed model for 0–50 in experiment 1. A Transfer learning model ImageNet
was used for existing and proposed models. The proposed model utilises an LSTM layer
which contains multiple Hyperparameters, including the rate of learning, the total number
of hidden layers, and the epoch durations, which helps the proposed model to reduce the
error rate and transfer learning aids to enhance the training time. We have also calculated
a ROC graph for 50 epochs, a graphic that displays how well an algorithm is for classifica-
tion across all the identification levels. Figure 10 presents the ROC graph of the proposed
Model for Waste classification; the graph was plotted based on two parameters True Posi-
tive Rate and False Positives Rate.
Similarly, Fig. 11 presents the Precision-Confidence Graph for all waste classes for the
proposed model. Figure 12 illustrates an F1-Measure graph for all waste classes. The graph
is plotted among the F1-score value and confidence results. A higher precision shows bet-
ter performance. The graph was plotted for various waste types, i.e., carboard-0, glass-1,
metal-2, plastics-3, and all classes-4.
Table 4 presents the error rate and total time comparison for the existing and proposed
Models for 0, 30, and 50 epochs. The proposed model takes 4589 s for 30 epochs, and
its error rate was also the lowest at 0.019578. Similar to 50 epochs, the error rate was
0.016578%, and the total time was 7586 s. A model with less error rate and time is always

Table 7  Precision, recall, Model Precision Recall F1-score Accuracy

F1-score, accuracy results
comparison for existing and
VGG-16 models 0.852 0.879 0.881 0.871
proposed model for 100 epochs
ResNet-34 0.874 0.887 0.892 0.881
ResNet-50 0.885 0.878 0.881 0.895
AlexNet 0.902 0.915 0.903 0.908
Proposed model 0.941 0.938 0.921 0.941

13
Multimedia Tools and Applications (2024) 83:29505–29529 29525

in demand; the proposed model achieved both factors. The proposed model takes 4589 s
for 30 epochs, and its error rate was also the lowest at 0.019578. Similar to the 50 epoch,
the error rate was 0.016578%, and the total time was 7586 s. A model with less error rate
and time is always in demand; the proposed model achieved both factors. Similarly, Table 5
presents an experimental result comparison of experiment 1 for existing and proposed
Models, i.e., Precision, Recall, F1-Score and Accuracy for 0–50 epochs and the [Link]
10 dataset. The proposed model achieves a precision of 95.4%, recall of 94.3%, F1-score
93.1% and accuracy of 95.7%, which is best compared to all the existing methods.
In experiment two on the same dataset, the splitting ratio was 70: 15: 15 (Training: Vali-
dation: Testing), and the number of Epochs was 0–50-100. Figure 13a presents training and
validation accuracy results for the proposed model from 0–50 epochs. When the numbers
of epochs are less, the training and validation accuracies are also less, but once the number
of epochs increases, the accuracy results are increased for training and validation from 75%
to 95.8%. A model with higher training and validation accuracy is always in demand, and
the proposed model achieved the same.
Figure 13b presents training and validation loss results for the proposed model from
0–50 epochs. Initially, the loss results are higher for epochs 0–20, but once we increase
it up to 50, the loss % is also decreased. Less loss results show better performance. Preci-
sion and recall results were calculated for 100 Epochs and dataset 70: 15: 15 (training:
validation: testing). Figures 14 and 15 present the proposed Model for Precision and Recall
experimental results. Graph 14 was plotted for precision Vs confidence, and Graph 15 for
recall Vs. Confidence. A model which generates precision is always in demand. The pro-
posed model achieved both. There are pros and cons to previous studies, and keeping them
in mind lets researchers develop a method that can effectively address all drawbacks. Previ-
ous studies have somewhat addressed the problem of garbage categorization and manage-
ment, but they all fall short in some way.
Some have proposed a hybrid strategy that combines elements of several different meth-
odologies. Using a combination of Deep Learning algorithms (CNN-LSTM and Transfer
Learning), the highest levels of accuracy have been attained. Over 94.1% accuracy was
attained when sorting garbage.
Many suggested systems, too, provide a "machine learning and deep learning" mix as
a solution [49–51]. The quest for a more precise and trustworthy system persists, never-
theless. The study’s overarching goal is to understand better and implement automatic
waste categorization systems, which aid in trash recycling. Various methods are available
for sorting trash, but many need human intervention. The government, the people, and the
business community may all benefit from implementing a completely autonomous system.
The ultimate goal of this study is to develop a low-cost, fully automated waste management
system that can efficiently classify garbage and return correct findings.

5 Conclusion and future works

For sustainable development of the environment, waste management is always in demand.

Multiple research works using deep learning methods presented various waste prediction
and classification research. This research proposed a hybrid waste recognition and clas-
sification model using CNN-LSTM with a Transfer learning feature. This research uti-
lised an online TrashNet dataset sample, with 24705 images separated into two classes of
organic waste (13880) and recyclable waste (10825) used to test the proposed model. An

13
29526 Multimedia Tools and Applications (2024) 83:29505–29529

experimental analysis was performed for two dataset divisions. In experiment one dataset
ratio of 80: 10: 10 (Training: Validation: Testing) and the number of epochs from 0–50, and
in experiment 2 dataset ratio of 70: 15: 15 (Training: Validation: Testing) and the number
of epochs from 0–100. Table 3 presents an experimental parameters overview for the pro-
posed model, and Fig. 8 shows the sample experimental results (Organic and Recyclable).
The proposed hybrid model and various existing CNN models (i.e., VGG-16,
ResNet-34, ResNet-50, and AlexNet) were implemented using Python and tested based
on performance measuring parameters, i.e., precision, recall, testing and training loss,
and accuracy. Each model was created with a range of epochs and an AME optimisation
algorithm. For the proposed method, the AME optimisation achieved the best optimisa-
tion and accuracy and the least modelling loss for training, validation, and testing. The
proposed model performed the highest precision of 95.45%, far better than the existing
deep-learning method. When putting this paradigm into action, for instance, in a smart city
initiative, there are many factors to consider. Autonomous vehicles equipped with a robotic
arm and high-definition camera, as well as efficient operating cycles throughout a wide
range of solid waste content areas, is one solution to these problems.
Finally, the proposed deep learning-based smart waste management and categorization
framework can completely transform current methods for handling waste. These structures
help contribute to a more sustainable and healthier future by lowering pollution levels, fos-
tering a sustainable environment, and increasing the efficiency with which waste is pro-
cessed. These systems are refined through continued advancements and research within
this discipline, leading to more successful methods for dealing with the ever-growing dif-
ficulties associated with waste management within a rapidly altering ecosystem.
In future research, we will explore variables such as operational socio-economic factors
and essential service expenditure that can affect waste generation. To create a waste man-
agement framework that can be better optimized and more economical additional research
may be carried out to determine the significant aspects. The proposed model also has one
limitation: it cannot sort the waste images based on the input; we will add this feature in
the future.

Author contributions Umesh Kumar and Sarita: Problem statement, the key concept, design and initial draft
preparation; Surjeet Dalal: Implementations, results analysis; Robertas Damaševičius: conceptualization
and methodology.

Data availability The dataset is openly available on the Kaggle online website. The dataset is available from
the corresponding author based on personal request.

Declarations
Conflict of interest The authors have no conflict of interest related to the research.

References
1. Alrayes FS, Asiri MM, Maashi MS, Nour MK, Rizwanullah M, Osman AE, Drar S, Zamani AS (2023)
Waste classification using vision transformer based on multilayer hybrid convolution neural network.
Urban Climate 49:101483
2. Wu T-W, Zhang H, Peng W, Lü F, He P-J (2023) Applications of convolutional neural networks for
intelligent waste identification and recycling: a review. Resour Conserv Recycl 190:106813
3. Li N, Chen Y (2023) Municipal solid waste classification and real-time detection using deep learning
methods. Urban Climate 49:101462

13
Multimedia Tools and Applications (2024) 83:29505–29529 29527

4. Zhou K, Sung-Kwun Oh, Pedrycz W, Qiu J (2023) Data preprocessing strategy in constructing con-
volutional neural network classifier based on constrained particle swarm optimisation with fuzzy
penalty function. Eng Appl Artif Intell 117:105580
5. Windrim L, Melkumyan A, Murphy RJ, Chlingaryan A, Leung R (2023) Unsupervised ore/
waste classification on open-cut mine faces using close-range hyperspectral data. Geosci Front
14(4):101562
6. Zhang H, Cao H, Zhou Y, Changle Gu, Li D (2023) Hybrid deep learning model for accurate clas-
sification of solid waste in the society. Urban Climate 49:101485
7. Lin K, Zhao Y, Kuo JH, Deng H, Cui F, Zhang Z, Zhang M, Zhao C, Gao X, Zhou T, Wang T
(2022) Toward smarter management and recovery of municipal solid waste: a critical review on
deep learning approaches. J Clean Prod 24:130943
8. Andeobu L, Wibowo S, Grandhi S (2022) Artificial intelligence applications for sustainable solid
waste management practices in Australia: a systematic review. Sci Total Environ 20:155389
9. Soundarya B, Parkavi K, Sharmila A, Kokiladevi R, Dharani M, Krishnaraj R (2022) CNN-based
smart bin for waste management. In: 2022 4th international conference on smart systems and inven-
tive technology (ICSSIT), IEEE, pp 1405–1409
10. Zhang H, Peeters J, Demeester E, Duflou JR, Kellens K (2022) A CNN-based fast picking method
for WEEE recycling. Procedia CIRP 1(106):264–269
11. Rubab S, Khan MM, Uddin F, Abbas Bangash Y, Taqvi SA (2022) A study on AI-based waste man-
agement strategies for the COVID-19 pandemic. ChemBioEng Reviews 9(2):212–226
12. Tiwari R, Dubey AK (2022) Development of Computer vision and deep learning based algorithm
to improve waste management system. In: 2022 2nd international conference on advance comput-
ing and innovative technologies in engineering (ICACITE), IEEE, pp 2178–2182
13. Patil M, Shaikh, N (2022) Waste classification using ANN, CNN and transfer learning. SSRN.
[Link]
14. Gothai E, Thamilselvan R, Natesan P, Keerthivasan M, Kabinesh K, Ruban DK (2022) Plastic
waste classification using CNN for supporting 3R’s principle. In: 2022 international conference on
computer communication and informatics (ICCCI), IEEE, pp 01–07
15. Bharti S, Fatma S, Kumar V (2022) AI in waste management: the savage of environment. Environ-
mental Informatics 97–123
16. Ihsanullah I, Alam G, Jamal A, Shaik F (2022) Recent advances in applications of artificial intel-
ligence in solid waste management: a review. Chemosphere 29:136631
17. Diqi M (2022) Waste classification using CNN algorithm. In: International conference on science
and technology innovation (ICoSTEC), vol 1, no 1, pp 130–135
18. Wang C, Qin J, Qu C, Ran X, Liu C, Chen B (2021) A smart municipal waste management system
based on deep-learning and Internet of Things. Waste Manage 1(135):20–29
19. Velis CA, Cook E, Cottom J (2021) Waste management needs a data revolution–Is plastic pollution
an opportunity? Waste Manage Res 39(9):1113–1115
20. Rajesh V, Rao KR, Devendra P, Babu EV, Venkatesh B, Nadipalli LS, Ahammad SH, Naidu TP
(2021) Waste segregation using CNN & IoT. NVEO-Natural Volatiles & Essential Oils Journal
8(5):4486–4494
21. Liang S, Gu Y (2021y) A deep convolutional neural network to simultaneously localise and recog-
nise waste types in images. Waste Manage 1(126):247–257
22. Simaiya S, Lilhore UK, Pandey H, Trivedi NK, Anand A, Sandhu J (2022) An improved deep neu-
ral network-based predictive model for traffic accident’s severity prediction. In: Ambient communi-
cations and computer systems. Springer, Singapore, pp 181–190
23. Erkinay Ozdemir M, Ali Z, Subeshan B, Asmatulu E (2021) Applying machine learning approach
in recycling. J Mater Cycles Waste Manage 23(3):855–871
24. Gondal AU, Sadiq MI, Ali T, Irfan M, Shaf A, Aamir M, Shoaib M, Glowacz A, Tadeusiewicz R,
Kantoch E (2021) Real time multipurpose smart waste classification model for efficient recycling in
smart cities using multilayer convolutional neural network and perceptron. Sensors 21(14):4916
25. Lilhore UK, Simaiya S, Kaur A, Prasad D, Khurana M, Verma DK, Hassan A (2021) Impact of
deep learning and machine learning in industry 4.0: impact of deep learning. In: Cyber-Physical,
IoT, and Autonomous Systems in Industry 4.0. CRC Press, pp 179–197
26. Trivedi NK, Simaiya S, Lilhore UK, Sharma SK (2021) COVID-19 pandemic: role of machine
learn-ing & deep learning methods in diagnosis. Int J Cur Res Rev 13(06):150–156
27. Sallang NC, Islam MT, Islam MS, Arshad H (2021) A CNN-based smart waste management sys-
tem using tensorflow lite and LoRa-GPS shield in internet of things environment. IEEE Access
15(9):153560–153574

13
29528 Multimedia Tools and Applications (2024) 83:29505–29529

28. Lilhore UK, Imoize AL, Lee CC, Simaiya S, Pani SK, Goyal N, Kumar A, Li CT (2022) Enhanced
convolutional neural network model for cassava leaf disease identification and classification. Math-
ematics 10(4):580
29. Lilhore UK, Simaiya S, Sandhu JK, Trivedi NK, Garg A, Moudgil A (2022) Deep learning-based pre-
dictive model for defect detection and classification in industry 4.0. In: 2022 international conference
on emerging smart computing and informatics (ESCI), IEEE, pp 1–5
30. Zheng H, Gu Y (2021) Encnn-upmws: Waste classification by a CNN ensemble using the UPM
weighting strategy. Electronics 10(4):427
31. Recycle Waste image dataset (organic, recyclable waste), Kaggle, online available at. [Link]
gle.​com/​techs​ash/​waste-​class​ifica​tion-​data. Access 9 Sep 2022
32. Alsabei A, Alsayed A, Alzahrani M, Al-Shareef S (2021) Waste classification by fine-tuning pre-
trained CNN and GAN. Int J Comput Sci Netw Secur 21(8):65–70
33. Abdallah M, Talib MA, Feroz S, Nasir Q, Abdalla H, Mahfood B (2020) Artificial intelligence applica-
tions in solid waste management: a systematic research review. Waste Manage 15(109):231–246
34. Kumar S, Yadav D, Gupta H, Verma OP, Ansari IA, Ahn CW (2020) A novel yolov3 algorithm-based
deep learning approach for waste segregation: towards smart waste management. Electronics 10(1):14
35. Nowakowski P, Pamuła T (2020) Application of deep learning object classifier to improve e-waste col-
lection planning. Waste Manage 15(109):1–9
36. Anh Khoa T, Phuc CH, Lam PD, Nhu LM, Trong NM, Phuong NT, Dung NV, Tan-Y N, Nguyen
HN, Duc DN (2020) Waste management system using IoT-based machine learning in university. Wirel
Commun Mob Comput 27:2020
37. Gyawali D, Regmi A, Shakya A, Gautam A, Shrestha S (2020) Comparative analysis of multiple deep
CNN models for waste classification. arXiv preprint arXiv:2004.02168
38. Abeygunawardhana AG, Shalinda RM, Bandara WH, Anesta WD, Kasthurirathna D, Abeysiri L
(2020) AI-driven smart bin for waste management. In: 2020 2nd international conference on advance-
ments in computing (ICAC), IEEE, vol 1, pp 482–487
39. Jahanbakhshi A, Momeny M, Mahmoudi M, Zhang YD (2020) Classification of sour lemons based on
apparent defects using stochastic pooling mechanism in deep convolutional neural networks. Sci Hortic
15(263):109133
40. Franchitti E, Pascale E, Fea E, Anedda E, Traversi D (2020) Methods for bioaerosol characterisation:
limits and perspectives for human health risk assessment in organic waste treatment. Atmosphere
11(5):452
41. Sidharth R, Rohit P, Vishagan S, Karthika R, Ganesan M (2020) Deep learning based smart garbage
classifier for effective waste management. In: 2020 5th International Conference on Communication
and Electronics Systems (ICCES), IEEE, pp 1086–1089
42. Hussain A, Draz U, Ali T, Tariq S, Irfan M, Glowacz A, Antonino Daviu JA, Yasin S, Rahman S
(2020) Waste management and prediction of air pollutants using IoT and machine learning approach.
Energies 13(15):3930
43. Bobulski J, Kubanek M (2019) CNN use for plastic garbage classification method. In: 25th ACM
SIGKDD conference on knowledge discovery and data mining, 4-8 August 2019, Anchorage, Alaska,
USA. ACM
44. Sunny MS, Dipta DR, Hossain S, Faruque HM, Hossain E (2019) Design of a convolutional neural
network based smart waste disposal system. In: 2019 1st international conference on advances in sci-
ence, engineering and robotics technology (ICASERT), IEEE, pp 1–5
45. Kumar NM, Mohammed MA, Abdulkareem KH, Damasevicius R, Mostafa SA, Maashi MS, Chopra
SS (2021) Artificial intelligence-based solution for sorting COVID related medical waste streams
and supporting data-driven decisions for smart circular economy practice. Process Saf Environ Prot
152:482–494. [Link]
46. Uzma Al-Obeidat F, Tubaishat A, Shah B, Halim Z (2020) Gene encoder: a feature selection technique
through unsupervised deep learning-based clustering for large gene expression data. Neural Comput
Appl 34:8309–8331. [Link]
47. Ullah S, Halim Z (2021) Imagined character recognition through EEG signals using deep convolu-
tional neural network. Med Biol Eng Compu 59(5):1167–1183
48. Wang Qi, Liu Z, Zhang T, Alasmary H, Waqas M, Halim Z, Li Y (2023) Deep convolutional cross-
connected kernel mapping support vector machine based on SelectDropout. Inf Sci 626:694–709
49. Bao N, Zhang T, Huang R, Biswal S, Su J, Wang Y, ... Cha Y (2023) A deep transfer learning net-
work for structural condition identification with limited real-world training data. Structural Control and
Health Monitoring, 8899806. [Link]

13
Multimedia Tools and Applications (2024) 83:29505–29529 29529

50. Zhao F, Wu H, Zhu S, Zeng H, Zhao Z, Yang X, ... Zhang S (2023) Material stock analysis of urban
road from nighttime light data based on a bottom-up approach. Environ Res 228:115902. [Link]
org/​10.​1016/j.​envres.​2023.​115902
51. Yu D, Guo J, Meng J, Sun T (2023) Biofuel production by hydro-thermal liquefaction of munici-
pal solid waste: process characterization and optimization. Chemosphere 138606. [Link]
1016/j.​chemo​sphere.​2023.​138606

Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and
institutional affiliations.

Springer Nature or its licensor (e.g. a society or other partner) holds exclusive rights to this article under
a publishing agreement with the author(s) or other rightsholder(s); author self-archiving of the accepted
manuscript version of this article is solely governed by the terms of such publishing agreement and applicable
law.

Authors and Affiliations

Umesh Kumar Lilhore1 · Sarita Simaiya2 · Surjeet Dalal3 · Robertas Damaševičius4

* Robertas Damaševičius
[Link]@[Link]
1
Department of Computer Science and Engineering, Chandigarh University, Mohali, Punjab, India
2
Department of Computer Science and Engineering, APEX Institute of Technology, Chandigarh
University, Mohali, Punjab, India
3
Department of Computer Science and Engineering, Amity University, Gurgaon, Haryana, India
4
Department of Applied Informatics, Vytautas Magnus University, Kaunas, Lithuania

13