Leveraging Transfer Learning for Plant Identification with Limited Data : A MobileNet-V2 Based

Approach

Himanshu Gupta Himanshu Kumar Diwedi Esha Tripathi

Pranveer Singh Institute of Technology Pranveer Singh Institute of Technology Pranveer

Singh Institute of Technology

Dr. APJ Abdul Kalam Technical University Dr. APJ Abdul Kalam
Technical University

Abstract

Plants have been an important pillar for the survival of humans from the existence of humans on earth
to now and further will be an integral part of our human existence.

Accurate plant classification is essential for various domains, including agriculture, biodiversity
conservation, and ecological research. However, developing robust models often proves challenging
due to the scarcity of labeled plant image datasets. To address this limitation, we propose a transfer
learning approach that harnesses the knowledge of a pre-trained MobileNetV2 model to effectively
classify a dataset of 5000 plant images representing 74 unique categories

Keywords

Transfer Learning , ConvoNet, ResNet, MobileNet

1. Introduction

This Paper is about the identification of Flora species that exist in our surroundings by the application of
Transfer Learning methodologies, which can provide a precise and fast result with good accuracy.

In the context of India's abundant and varied plant life, precise plant identification holds significant importance
for ecological research, the surveillance of endangered species, and the encouragement of citizen science
projects. Nevertheless, prevailing plant identification techniques face challenges when dealing with the distinct
attributes of Indian plant species and the scarcity of well-annotated datasets. This research aims to fill this void
by creating a specialized transfer learning model tailored to ensure the precise identification of flora species
indigenous to India.

2. Ease of Use

Democratizing Plant Identification in India: This plant identification model prioritizes user experience
through several accessibility features. Its intuitive interface caters to users of diverse technical backgrounds,
allowing seamless navigation and effortless application. The visually appealing and straightforward design
fosters accessibility for both researchers and citizen scientists.
Simplifying the Identification Process: The model boasts a streamlined workflow, minimizing identification
complexity. Clear instructions and prompts guide users through each step, from uploading images to interpreting
accurate results. This emphasis on simplicity facilitates efficient use by experts and encourages participation
from those with varying botanical knowledge.

Accessibility Across Devices: Adaptability to different devices and platforms further increases ease of use.
Whether accessed on a desktop, tablet, or smartphone, users can seamlessly engage with the identification tool,
promoting widespread adoption and participation in ecological studies and citizen science initiatives.

Empowering Everyone to Contribute: By prioritizing a user-centric approach, this model aims to democratize
plant identification in India. It caters to a broad audience, enabling anyone interested in ecological research and
conservation efforts to contribute actively.

2.Prior Works

Interactive plant identification based on social image data[1]

According to Joly et al.The main contribution of this paper is to provide an innovative collaborative workflow
focused on image-based plant identification as a mean to enlist new contributors and facilitate access to
botanical data. The major contribution of this paper was in the field of making efforts more successful toward
the real-world image-based identification problem. This Research paper has addressed the major issue of the
"one-Shot-Problem" which states that the dataset that is available nowadays on the web do have not so plenty of
images in each of the classes in the dataset, involving the fact that many of datasets have no, image for a class
that has been mentioned in labels of the dataset. This leads to ambiguity since the model when exposed to
classes that were absent in the training set leads to the downfall of the fine-tuning performance.The paper also
tries to impart the light to address the issue of "Heterogeneity Problem"[1].

LifeCLEF Plant Identification Task [2]

This paper aims to evaluate the pre-existing plant identification systems at very large scale, which approximates
closely to real-world conditions of a real-world biodiversity scenarios. This research paper implies to use
"PlantCLEF 2015" Dataset which is composed of 113,205 images belonging to 41,794 observations of 1000
species of all Flora species existing in Western European regions. Under this research the whole PlantCLF
dataset was split into two parts as usual, the main difference was made in a testing set that was consisting of 1/2
of observations of each species with constrained random rule where as the remaining observatoins were kept in
the reference training set.

Plant Identification in an open-world(LifeCLEF 2016)[3]

The paper evaluated system performance using a metric called "mAP-open," which assesses how well the
system handles open-set recognition tasks. This means the system needs to be robust against encountering
unknown or unseen categories.

The key takeaway was that, compared to the previous year, Convolutional Neural Networks (CNNs) displayed a
natural level of robustness against unknown categories in the test set. However, none of the additional rejection
methods employed by participants further improved this inherent robustness of CNNs.
Feature Extraction and Automatic Recognition of Plant Leaf Using Artificial Neural Network [4]

The research aims to automate plant leaf recognition by extracting visual features like shape, margin, dent, and
vein from leaf images. By implementing an artificial neural network approach, the study demonstrates the
effectiveness and superiority of this method through experimental results. The system achieves rapid recognition
times and high accuracy in distinguishing various types of leaves. Previous methods focused mainly on contour
features, neglecting other important aspects like leaf dent and vein. This study fills that gap by analyzing and
extracting domain-related visual features specific to plant leaves. The results validate the system's ability to
accurately identify plant leaves, showcasing the potential for advanced automated plant recognition systems.
This research contributes to the field by providing a detailed analysis of leaf features and presenting a practical
approach for efficient and accurate plant leaf recognition.

Deep Learning for Plant Identification in Natural Environment [5].

In this study, a deep learning model called ResNet26 with 26 layers and 8 residual building blocks for plant
image classification. The model successfully achieved recognition accuracy of 91.78\% on the BJFU100 dataset,
demonstrating its effectiveness in this task., collected from natural scenes using mobile devices. The ResNet26
outperformed other reference models and conventional classification methods. The deep learning approach
simplifies data preprocessing compared to traditional methods. Experiments on the Flavia leaf dataset showed
the effectiveness of the ResNet26 model, with test accuracies of 99.44\%, 98.95\%, and 98.60\% for ResNet18,
ResNet34, and ResNet50, respectively

Machine Learning for Plant Species Classification using Leaf Vein Morphometric [6]

The sheer number of plant species demands an alternative to manual identification, leading to the development
of this automated classification system using machine learning. Leveraging leaf features like shape, color,
texture, and vein patterns, the system tackles this challenge with robust data pre-processing techniques. Images
are transformed into numerical data via contour analysis, paving the way for efficient feature extraction.

The Flavia dataset, boasting 1907 images representing 32 diverse species, serves as the experimentation ground.
Veins, crucial identifiers, are extracted using the innovative power of genetic algorithms and Ant Colony
optimization. Three state-of-the-art classifiers – Support Vector Machines, k-Nearest Neighbours, and Random
Forests – tackle the identification challenge. SVM's prowess in handling high-dimensional data and kNN's
collaborative approach through majority voting equip the system with exceptional accuracy.

Ultimately, this proposed system aspires to be a valuable tool for botanists, enabling swift and accurate plant
species identification. By empowering efficient taxonomic efforts, the system contributes to the vital mission of
preserving our precious biodiversity.

Plant Recognition Using Morphological Feature Extraction and Transfer Learning over SVM and
AdaBoost [7]
The article proposes a plant species recognition model based on morphological features extracted from leaf
images using SVM with adaptive boosting. It emphasizes the significance of leaf-based features for AI-based
models due to their easy accessibility. The model involves pre-processing, feature extraction, and classification.
Various morphological features like centroid, axis lengths, perimeter, and orientation are extracted from leaf
images.

Literature review shows successful use of SVM and boosting techniques for plant recognition. SVM is a
memory-efficient classification model, while boosting combines weak classifiers for a strong one. Feature
selection is crucial for efficient convergence. The study aims to achieve high accuracy in plant species
identification through morphological features and adaptive boosting.

Leafsnap: A ComputerVision System for Automatic Plant Species Identification [8] Leafsnap; significantly
contributes to the realm of plant species identification by providing a cutting-edge solution that
bridgestechnology and botany. Its user-friendly interface and accurate identification capabilities make it a
valuable tool for both experts and non-experts, fostering a deeper understanding of plant biodiversity.

The research paper introduces Leafsnap, a pioneering computer vision system designed for the automatic
identification of plant species. The central idea revolves around leveraging advanced image recognition
techniques to create an efficient and accurate tool for botanists, researchers, and nature enthusiasts.

The primary objective is to develop a robust system capable of automatically identifying plant species based on
leaf images.

Addressing the growing need for technology-driven solutions in the field of botany and plant taxonomy.

The authors curate a substantial dataset comprising high-quality leaf images representing a diverse range of
plant species.

Employing state-of-the-art image recognition techniques, the system

extracts key features from leaves to enable accurate species identification.

Leaf Snap goes beyond traditional research tools by translating its capabilities into a user-friendly mobile
3application.

The mobile app allows users to capture leaf images using their

smartphones and receive instant species identification, promoting accessibility and engagement.

Addressing challenges related to variations in leaf appearance,

environmental conditions, and potential overlapping features among species.

The paper discusses the algorithms and methodologies implemented to mitigate these challenges and enhance
the accuracy of species identification.

Pre-trained deep learning-based classification of jujube fruits according to their maturity level. Neural
Computation and Application [9] The study compared two convolutional neural network paradigms, AlexNet
and VGG16, for classifying jujube fruits based on maturity levels. The models achieved the best accuracy of
94.17\% and 97.65\% for actual and augmented images, respectively. The improved success rate of VGG16
makes it recommended for automated harvesting and sorting of jujube fruits.

Contemporary machine learning applications in agriculture: Quo Vadis. Concurrency and Computation
Practice and Experience.[10] The study suggests that machine learning can improve agricultural automation
for cost-effectiveness, efficiency, and precision. It serves as a guide for researchers and manufacturers.

Maturity Grading of Jujube for Industrial Applications Harnessing Deep Learning. 10.21203.[11] The
research compares a customised CNN model and two classical models, VGG16 and AlexNet, for grading jujube
fruits based on maturity stages. The customised CNN model achieved high grading accuracy of 99.44% and
97.53\%, with reduced computation time and training parameters. The model outperformed both classical
models, demonstrating its potential in jujube fruit classification.

Research Review and Literature Perception Towards Medicinal Plants Classification Using Deep
Learning Techniques[12] There are several ways to improve the powerful medicinal plant classifier that can be
useful in real-time classification. This work uses several reliable and effective deep-learning techniques for
classifying plants based on leaf images

Pre-trained deep learning-based classification of jujube fruits according to their maturity level. Neural
Comput \& Applic 34, 13925 [13] The manual classification of crops based on their maturity levels for
harvesting and packaging purpose is a tedious process. However, the emergence of machine learning/deep
learning techniques has opened up the ways in this direction, but its practical success is still limited. In this
research, we examined two convolutional neural network paradigms (i.e., AlexNet and VGG16) utilizing a
transfer-learning approach for classifying the jujube fruits based on their maturity level (i.e., unripe, ripe, and
over-ripe).

CNN-based medicinal plant identification and classification using optimized SVM. Multimed Tools Appl
(2023) [14]. An Enhanced Convolutional Neural Network architecture (using modified ResNet50) with
Progressive Transfer Learning (ECNN-PTL) has been proposed in this paper. The suggested method uses an
improved ReNet50 framework for feature extraction along with PTL. Classification has been done using an
Optimized Support Vector Machine (OSVM).

Proposed Methodology

The proposed system tries to make use of the concept of Transfer Learning which states that when it is difficult
to model a problem and there is less data needed to make a good model, it is preferable to use some model that
has been trained over large datasets and make customization on the same pre-trained model as per the
requirement to design a solution to the problem. Fig. 1 depicts the flowachart of proposed methodology.

For this problem, we have implied to use the MobileNetV2 model which has been trained over a large size
dataset called ImageNet which contains approx. 1 lakh images in the training set and 50,000 in the validation
set.
In our case, we have a dataset of 5000 images belonging to 74 unique different categories of plants. The dataset
has been split in the ratio of 80:20 where the training set contains 3700 images and the test set contains a total
of 1500 images.

Transfer Learning Framework: We employed MobileNetV2, a pre-trained convolutional neural network

(CNN) known for its efficiency and performance, as the foundation for our model.

Model Customization: We adapt the pre-trained model to our specific plant classification task by fine-tuning
its final layers to accommodate the 74 target categories.

Dataset Preparation

 We meticulously curate a dataset comprising 4200 plant images across 74 distinct classes.
 To ensure a comprehensive evaluation, we split the dataset into an 80\% training set (3700 images) and
a 20\% test set (1500 images).

6. Tables

Model Accurac Precisio Recall F Score

y n
CNN 256N 0.6 0.42 0.26 0.10
ResNet 50M 0.54 0.231 0.35 0.45
REsNet with FCN 0.3 0.12 0.36 0.2
VGG 16 IMP 245 0.016 0.0012 0.0010 0.008
MobileNet 0.68 0.15 0.435 0.196
MobileNet wit TB 0.814 0.65 0.75 0.792
MobileNet with Early 0.729 0.63 0.43 0.345
Stopping
TABLE I
EXPERIMENTAL TABLE FOR COMPARATIVE PERFORMANCE ANALYSIS
OF DIFFERENT METHODS ON OUR DATASET

No Plant Name No. of

. images
0 Amaranthus Green 123
1 Amaranthus Red 89
2 Asthma Plant 81
3 Avaram 66
4 Balloon vine 246
5 Bellyache bush 81
(Green)
6 Benghal dayflower 58
7 Betel Leaves 127
8 Big Caltrops 64
9 Black Night Shade 108
10 Black-Honey Shrub 105
11 Bristly Wild Grape 66
12 Butterfly Pea 51
13 Cape Gooseberry 122
14 Celery 82
15 Chinese Spinach 60
16 Common Wireweed 60
17 Coriander Leaves 60
18 Country Mallow 60
19 Crown flower 65
20 Curry Leaf 65
21 Dwarf Copperleaf 88
(Green)
22 Dwarf copperleaf 79
(Red)
23 False Amarnath 101
24 Fenugreek Leaves 80
25 Giant Pigweed 103
26 Gongura 53
27 Green Chireta 69
28 Holy Basil 54
29 Indian CopperLeaf 124
30 Indian Jujube 66
31 Indian Sarsaparilla 84
32 Indian Stinging Nettle 106
33 Indian Thornapple 57
34 Indian pennywort 64
35 Indian wormwood 67
36 Ivy Gourd 51
37 Kokilaksha 59
38 Lagos Spinach 84
39 Lambs Quarters 69
40 Land Caltrops (Bindii) 63
41 Lettuce Tree 64
42 Madagascar 55
Periwinkle
43 Madras Pea Pumpkin 53
44 Malabar Catmint 65
45 Malabar Spinach 106
(Green)
46 Mexican Mint 105
47 Mexican Prickly 55
Poppy
48 Mint Leaves 125
49 Mountain Knotgrass 78
50 Mustard 80
51 Nalta Jute 73
52 Night blooming 54
Cereus
53 Palak 84
54 Panicled Foldwing 68
55 Prickly Chaff Flower 54
56 Punarnava 76
57 Purple Fruited Pea 70
Eggplant
58 Purple Tephrosia 59
59 Rosary Pea 66
60 Shaggy button weed 65
61 Siru Keerai 68
62 Small Water Clover 57
63 Spiderwisp 55
64 Square Stalked Vine 159
65 Stinking 66
Passionflower
66 Sweet Basil 123
67 Sweet flag 75
68 Tinnevelly Senna 53
69 Trellis Vine 57
70 Velvet bean 72
71 Water Spinach 55
72 coatbuttons 85
73 heart-leaved moonseed 192
TABLE II IMAGE CLASSES DITRIBUTION
TABLE III TESTING PERFORMANCE

Parameters Testing
Performance
Accuracy 87
Validation 86
Accuracy
Precision 88.26
Recall 86.88
7. Figures
Figure I. Flowchart for proposed Methodology

Figure II. Dataset Collection Phase

Fig. III snapshot of Whole Datset

Fig 4. Confusion Matrix for Model Performance

7. Results

Our proposed transfer learning approach based on MobileNetV2 achieved an impressive accuracy of 86.0\% on
the test set, demonstrating its effectiveness in plant classification even with a limited dataset. This performance
is further supported by a validation accuracy of 88.0\%, indicating good generalizability of the model.

Further analysis revealed a precision of 88.88\% and a recall of 86.31\%, suggesting a well-balanced trade-off
between correctly identifying positive and negative instances. The F1-score of 86.81\% validates the overall
robustness of the model in both precision and recall. These results are encouraging, considering the challenging
nature of plant classification with relatively small datasets.

8. Conclusion

Our proposed plant identification system successfully integrates transfer learning with a privacy-preserving
design, demonstrating its potential for accurate and efficient plant classification. The initial results, with an
accuracy of 88\% and balanced precision and recall metrics, are highly encouraging. However, further testing
and refinement are necessary to optimize performance and user experience for real-world deployment.
Additionally, incorporating larger and more diverse training datasets and collecting user feedback will
continually improve the system's accuracy and adaptability. By pursuing these avenues, we can ensure this
application remains a valuable tool for users of all experience levels to explore and learn about the fascinating
world of plants.

9. References

[1] A. Joly, H. Go¨eau, P. Bonnet et al., “Interactive plant identifica

tion based on social image data,” Ecological Informatics,vol.23,

pp.22–34,2014.

[2] H.Go¨eau, P. Bonnet, andA.Joly, “LifeCLEF plant identification

task 2015,” in Proceedings of the Conference and Labs of the

Evaluation Forum (CLEF ’15),2015..

[3] H. Go¨eau, P. Bonnet, and A. Joly, “Plant identification in an

open-world (lifeclef 2016),” in Proceedings of the CLEF working

notes,vol.2016,2016.

[4] Feature Extraction and Automatic Recognition

of Plant Leaf Using Artificial Neural Network

Qingfeng Wu, Changle Zhou and Chaonan Wang.

[5] Deep Learning for Plant Identification in Natural Environment

YuSun,YuanLiu,GuanWang,andHaiyanZhan.

[6] Machine Learning for Plant Species Classification using Leaf Vein Morphometric

[7] Plant Recognition Using Morphological Feature Extraction andTransfer Learning over SVM and
AdaBoost

ShubhamMahajan 1,*, Akshay Raina 2

[8] Leafsnap: A Computer Vision System for Automatic Plant Species Identification

[9] Mahmood, A., Singh, S.K. and Tiwari, A.K. Pre-trained deep learning-based classification of
jujube fruits according to their maturity level. Neural Comput and Applic 34, 13925–13935 (2022).

[10] Mahmood, Atif and Tiwari, Amod and Singh, Sanjay and Udmale, Sandeep. (2022).
Contemporary machine learning applications in agriculture: Quo Vadis. Concurrency and
Computation Practice and Experience. 34. e6940. 10.1002cpe.6940.
[11] Mahmood, Atif and Tiwari, Amod and Singh, Sanjay and Ali, Mohsin. (2023). Maturity Grading
of Jujube for Industrial Applications Harnessing Deep Learning. 10.21203/rs.3.rs-2561485v1.

[12] Diwedi, H.K., Misra, A., Tiwari, A.K., Mahmood, A. (2023). Research Review and Literature
Perception Towards Medicinal Plants Classification Using Deep Learning Techniques. In: Borah, S.,
Gandhi, T.K., Piuri, V. (eds) Advanced Computational and Communication Paradigms . ICACCP
2023. Lecture Notes in Networks and Systems, vol 535. Springer, Singapore.

[13] Mahmood, A., Singh, S.K. and Tiwari, A.K. Pre-trained deep learning-based classification of
jujube fruits according to their maturity level. Neural Comput \& Applic 34, 13925–13935 (2022).

[14] Diwedi, H.K., Misra, A. and Tiwari, A.K. CNN-based medicinal plant identification and
classification using optimized SVM. Multimed Tools Appl (2023).