Supply Chain Optimization Using Machine Learning-Based Shelf Life

Prediction for Perishable Produce

Report submitted to

Indian Institute of Technology, Kharagpur

for the award of the degree

of

Master of Technology
In
Operations Research and Data Analytics
by

Vivek Navnath Godase

(24IM61R07)

Under the guidance of

Prof. V.N.A. Naikan

DEPARTMENT OF INDUSTRIAL AND SYSTEMS ENGINEERING

INDIAN INSTITUTE OF TECHNOLOGY, KHARAGPUR
April 2025
 DECLARATION

I certify that

a. The work contained in this report is original and has been done by me under the guidance
of my supervisor.
b. The work has not been submitted to any other Institute for any degree or diploma.
c. I have followed the guidelines provided by the Institute in preparing the report.
d. I have conformed to the norms and guidelines given in the Ethical Code of Conduct of
the Institute.
e. Whenever I have used materials (data, theoretical analysis, figures, and text) from other
sources, I have given due credit to them by citing them in the text of the report and
giving their details in the references. Further, I have taken permission from the
copyright owners of the sources, whenever necessary.

Signature of the Student

 CERTIFICATE

This is to certify that the Dissertation Report entitled Supply Chain Optimization Using Machine Learning-
Based ShelfLife Prediction for Perishable Produce submitted by Mr. Vivek Navnath Godase to Indian
Institute of Technology, Kharagpur, India, is a record of bonafide Project work carried out by his under my
supervision and guidance and is worthy of consideration for the award of the degree of Master of Technology
in Industrial Engineering and Management of the Institute.

Date: Supervisor
 TABLE OF CONTENTS

S.No. Title
1. Introduction
2. Literature Review
3. Research Gap

3. Methodology:
1. Data Collection and Preprocessing
2. Data Augmentation
3. Model Development:
a) Custom CNN
b) InceptionV3 Transfer Learning Model
4. Model Training & Evaluation

4. References
INTRODUCTION

Perishable produce such as fruits and vegetables constitute a vital segment of the food supply chain
but is highly vulnerable to spoilage due to its limited shelf life. The rapid deterioration of these
goods, especially under suboptimal storage and transportation conditions, leads to substantial losses
at various points in the supply chain. These inefficiencies not only impact the profitability of
businesses but also contribute to increased food waste, environmental degradation, and resource
misallocation.

One of the primary reasons for such losses is the inability of traditional supply chain systems to
accurately assess the remaining shelf life of produce in real time. Current practices largely depend
on generalized estimates or visual inspections, which often fail to reflect the true quality and
freshness of the items. Consequently, decisions regarding distribution, inventory management, and
delivery routing are made without adequate insight into the actual condition of the products,
resulting in avoidable spoilage and customer dissatisfaction.

The emergence of machine learning technologies offers a powerful solution to this challenge. By
utilizing image-based data alongside environmental parameters such as temperature and humidity,
machine learning models can predict the remaining shelf life of perishable goods with greater
accuracy. These predictions, when embedded within supply chain management systems, can be used
to make proactive and data-driven decisions, ensuring that fresher items reach consumers more
efficiently while minimizing waste.

This project proposes an integrated framework that combines machine learning-based shelf life
prediction with supply chain optimization strategies. The approach involves the development of
predictive models trained on visual and contextual data to estimate freshness levels, followed by the
implementation of optimization algorithms to manage logistics, routing, and inventory based on
these predictions.

The aim is to create a smarter, more responsive supply chain that can dynamically adapt to the real-
time status of perishable produce. This methodology is expected to result in significant
improvements in supply chain performance, including reduced spoilage rates, improved delivery
timing, and enhanced resource utilization. Ultimately, this work seeks to contribute to a more
sustainable and efficient agricultural supply chain, aligned with the goals of food security and
environmental stewardship.
LITERATURE REVIEW
Varsha Bhole et al. (2021) food wastage due to the perishability of fruits remains a critical
global issue, necessitating accurate and non-destructive shelf-life prediction methods to
improve supply chain efficiency. Traditional techniques like chemical analysis are destructive
and time-intensive, prompting a shift toward AI-based, non-invasive solutions such as
computer vision, thermal imaging, hyperspectral imaging, and Near-Infrared Spectroscopy
(NIRS). While earlier methods using RGB imaging and machine learning showed limited
accuracy, recent advances leverage thermal imaging combined with deep learning and transfer
learning to assess intrinsic fruit qualities like bruising and temperature gradients. Lightweight
pre-trained CNN architectures—such as SqueezeNet, MobileNetV2, and ShuffleNet—have
been effectively adapted for classifying mango shelf life into 19 categories, with ShuffleNet
achieving an impressive 98.15% accuracy using thermal images. Compared to previous models
like fuzzy logic, Gaussian Mixture Models (GMM), or VIS/NIR spectroscopy, this approach
demonstrates superior generalizability and precision. Despite these advancements, challenges
remain in extending the method to various fruit types, enhancing real-time applicability,
enriching datasets under diverse environmental conditions, and integrating predictive tools into
end-to-end supply chain systems.

Yuan and Chen (2024) proposed a deep learning-based approach for detecting the freshness
of fruits and vegetables using image data. They utilized a dataset of 12,000 labeled images
across 20 categories (10 fruits, 10 vegetables), applying three pre-trained CNNs (GoogLeNet,
DenseNet-201, ResNeXt-101) for feature extraction. The extracted deep features were fused
and reduced via Principal Component Analysis (PCA) to optimize dimensionality.
Classification was performed using SVM, LDA, and Bagging, with SVM achieving the highest
accuracy of 96.98%. The method showed strong generalization across varying lighting and
backgrounds, offering a scalable, non-destructive, and efficient solution for automated
freshness detection in food supply chains. However, performance varied for certain items,
indicating room for improvement in handling high intra-class variability.

Dawei Li et al.(2024) machine learning (ML) has revolutionized the prediction and regulation
of shelf life in fruits and vegetables by providing non-destructive, rapid, and highly accurate
alternatives to traditional methods, which are often labor-intensive, destructive, and subjective.
Traditional ML algorithms such as Support Vector Machines (SVM), Decision Trees (DT),
Random Forests (RF), k-Nearest Neighbors (k-NN), and Extreme Learning Machines (ELM)
have demonstrated strong performance in predicting spoilage stages and freshness levels using
sensory, environmental, and physicochemical data. Deep learning models, including Artificial
Neural Networks (ANN), Convolutional Neural Networks (CNN), Radial Basis Function
Neural Networks (RBFNN), and Generative Adversarial Networks (GAN), further enhance
accuracy by capturing complex patterns from visual, spectral, and biochemical inputs, enabling
tasks like ripeness classification and spoilage detection with accuracies up to 99%. ML models
have been effectively integrated with environmental data (e.g., temperature, humidity),
physiological indicators (e.g., pH, TSS), and advanced sensing technologies such as
hyperspectral imaging and electronic noses for real-time, non-invasive quality assessment.
Additionally, the integration of ML with metabolomics and transcriptomics has enabled the
identification of key biomarkers linked to freshness, supporting targeted shelf life extension
strategies. Despite these advances, challenges remain in optimizing model parameters,
expanding datasets, enhancing model explainability, and embedding ML into intelligent supply
chains. Future research is expected to focus on explainable AI, data augmentation, multi-model
fusion, and real-time integration with logistics systems to minimize food waste and enhance
sustainability across the food industry.

Sowmiya R. et al. (2024) address the pressing issue of fruit waste, with the FAO reporting that
approximately one-third of global food production is lost annually. Traditional visual
inspections of fruit quality are subjective, inconsistent, and inefficient, making them unsuitable
for large-scale operations. To overcome these limitations, recent studies have explored machine
learning (ML) and computer vision (CV) approaches. Convolutional Neural Networks (CNNs)
and Support Vector Machines (SVMs) have shown high accuracy in classifying fruit ripeness
and freshness based on visual features such as color, shape, and texture. More advanced models
like YOLOv3 and semantic segmentation are used to detect external and internal fruit defects.
To enhance prediction accuracy, environmental factors—such as temperature, humidity, and
ethylene levels—have been integrated with image-based features using regression-based ML
models. Sowmiya et al. propose a two-step framework that combines CNN-based freshness
classification with shelf-life prediction, developed using tools like TensorFlow and Keras.
Despite promising results, the model’s generalizability remains a challenge. The authors
recommend future research focus on expanding fruit datasets, using advanced neural
architectures (e.g., LSTM, GRU), enabling real-time deployment, and integrating the system
with supply chain platforms.

Dr. K. Vijay Baskar et al. (2024) the post-harvest loss of fruits and vegetables remains a
significant challenge, particularly in countries like India, which contributes about 14% to global
production. Traditional shelf-life prediction methods, relying on fixed temperature rules and
manual inspection, often lack precision and lead to substantial spoilage. Recent advancements
have seen a shift towards data-driven approaches, especially deep learning models utilizing
real-time environmental data such as temperature and CO₂ levels for predicting remaining shelf
life (RSL) and optimizing transport and storage. This review examines recent studies
employing convolutional neural networks (CNNs), transfer learning models (e.g., VGG-16,
MobileNetv2, AlexNet), and object detection frameworks like YOLOv5 and Faster R-CNN.
These models have demonstrated high accuracy—often exceeding 98%—in classifying
ripeness and estimating shelf life using RGB images, thermal imaging, and sensor data.
Notably, lightweight CNN architectures are effective for mobile deployments, and thermal and
CO₂ data enhance prediction reliability. However, current research shows gaps in scalable real-
time deployment, integration of environmental sensors with vision-based models, and model
adaptability to dynamic storage and transport conditions. While the integration of deep learning
with sensor data marks a major step toward intelligent food logistics, future work should focus
on developing robust, adaptive systems capable of real-time learning and deployment across
the supply chain.

Rashvand et al. (2025) proposed Accurate shelf-life prediction is essential in the food industry
to ensure safety, reduce waste, and optimize supply chains. Traditional methods such as
microbial, chemical, and sensory analyses are effective but time-consuming and unsuitable for
real-time applications. Artificial intelligence (AI), including machine learning (ML) and deep
learning (DL), offers data-driven, scalable solutions capable of modeling complex interactions
between biochemical, microbial, and environmental factors. Techniques like decision trees,
support vector machines, convolutional and recurrent neural networks, and transfer learning
have shown strong potential, especially when integrated with sensor technologies such as
spectroscopy, electronic noses, and machine vision. These AI systems enable non-destructive,
real-time monitoring of spoilage in fruits, vegetables, meat, dairy, and beverages. Despite
promising results, challenges remain, including data quality, model generalizability, and high
implementation costs. Future innovations in IoT, cloud computing, and smart packaging may
help overcome these barriers, making AI-based shelf-life prediction more accessible and
sustainable across the food supply chain.
RESEARCH GAP

Research Gaps in Shelf-Life Prediction for Food Products:

1. Model Generalization: Existing models often struggle with generalizing across varying
conditions (lighting, background, and fruit types) and produce performance inconsistencies
in real-world environments.
2. Integration of Multi-modal Data: While different sensor data (visual, spectral,
environmental) show promise, their effective integration for more accurate shelf-life
prediction remains underexplored.
3. Real-time and Scalable Deployment: Challenges exist in deploying AI-based systems in
real-time, large-scale supply chains, especially in dynamic environments that require real-
time learning and adaptation.
4. Data Quality and Availability: The lack of high-quality labeled datasets limits model
performance. Current models struggle under varying environmental conditions, highlighting
the need for more diverse data.
5. Cost and Accessibility: High costs of advanced technologies like hyperspectral imaging
limit accessibility, especially for small-scale producers and retailers.
6. Explainability and Interpretability: The black-box nature of deep learning models makes
them difficult to interpret, posing a challenge for their practical deployment in decision-
making processes.
7. Integration with Logistics Systems: Current models have yet to integrate effectively with
real-time logistics data (e.g., temperature, humidity), which is crucial for accurate shelf-life
prediction across the supply chain.

PROBLEM STATEMENT
Managing the supply chain for perishable agricultural items like fruits and vegetables is challenging
due to their limited and variable shelf life. Inefficient handling, lack of real-time freshness tracking,
and delayed logistics decisions often lead to significant spoilage and economic losses. Traditional
systems lack the ability to assess product quality dynamically during storage and distribution.

This project aims to build a machine learning-based system that predicts the remaining shelf life of
perishable produce using image data. These predictions are used to optimize supply chain decisions
such as sorting, routing, and inventory turnover. By relying solely on visual cues captured from the
produce at different stages, the proposed approach enhances freshness monitoring without the need
for costly or complex sensor infrastructure.
METHODOLOGY
1. Data Collection and Preprocessing
 This dataset contains information about arrival time of fruits (apples and bananas) to

market and there respective shelf-life duration.

 Images of bananas were collected and categorized into stages: Apple(1-5), Apple(6-10),
Apple(11-15), Apple(16-20), Apple Expired ; Banana(1-2) days, Banana (3-4) days
Banana (5-7) days, Banana(8-10) days, Banana(Expired)
 All images resized to a standard shape (100×100 for basic model, 224×224 for InceptionV3).
 Normalization: Pixel values scaled to the range [0, 1].
 Labels encoded using one-hot encoding.

2. Data Augmentation
Techniques used:
 Horizontal flipping
 Random rotation (90°, 180°, 270°)
 Zoom, shear, and shift transformations
This step improved model generalization and mitigated overfitting.

3.Model Development
Two models were developed:

a. Custom CNN Model

 Layers:

o Convolutional + MaxPooling layers

o Flatten + Dense layers

Fig. CNN Architecture

 Accuracy: Achieved ~80.6% on the test set.

b. InceptionV3 Transfer Learning Model

 Base: InceptionV3 (pretrained on ImageNet)
 Fig Inception V3 CNN Architecture

 Modifications:
o GlobalAveragePooling2D

o Dense layers with ReLU and softmax

o Dropout layer to reduce overfitting

 Training Strategy:

o Initial training with frozen base

o Fine-tuning of top 50 layers with a reduced learning rate

4. Model Training & Evaluation

 Epochs: 32 (basic) + 10 + 20 (InceptionV3)
 Loss Function: Categorical Crossentropy
 Optimizer: Adam
 Callbacks: EarlyStopping, ModelCheckpoint
5. Output results:

Conclusion:

Deep learning, especially transfer learning with InceptionV3, can automatically classify the shelf
life stage of bananas, apple using images. This solution can be integrated into smart sorting systems
or quality control applications in the fruit supply chain, helping to reduce waste and ensure product
quality.

Limitation:
A key limitation of this model is its reliance solely on image data, which cannot capture internal
factors like texture or moisture that also influence Apple, banana shelf life. Visually similar stages
(e.g., day 5–7 vs. 8–10) often lead to classification errors due to overlapping features. Additionally,
the dataset is relatively small and may lack diversity in lighting, background, and angles, affecting
generalization. High computational requirements and potential human bias during manual labeling
also pose practical challenges for large-scale or real-time deployment.
REFERENCES

• https://www.researchgate.net/publication/351036890_A_Transfer_Learning-
based_Approach_to_Predict_the_Shelf_life_of_Fruit
• https://www.sciencedirect.com/science/article/pii/S2665927123002241?via%3Dihub
• https://www.ijraset.com/research-paper/predicting-shelf-life-of-fruits-a-multifactorial-
approach-using-intrinsic
• https://www.researchgate.net/publication/384319189_Research_Progress_of_Machine_Le
arning_in_Extending_and_Regulating_the_Shelf_Life_of_Fruits_and_Vegetables
• https://dialnet.unirioja.es/servlet/articulo?codigo=9906284
• https://www.sciencedirect.com/science/article/pii/S0924224425001256?via%3Dihub