PaperlayerView exact revision on arXiv

Source-linked AI summary

MangoLeafBD: A Comprehensive Image Dataset to Classify Diseased and Healthy Mango Leaves

Sarder Iftekhar Ahmed, Muhammad Ibrahim, Md. Nadim, Md. Mizanur Rahman, Maria Mehjabin Shejunti, Taskeed Jabid, Md. Sawkat Ali

arXiv:2209.02377v1cs.CVcs.LG

TL;DR

Agricultural machine-learning research lacks representative, publicly available datasets. This paper develops a ready-to-use mango-leaf dataset from Bangladesh, comprising 4,000 images of about 1,800 distinct leaves across seven diseases.

  • Problem

    Standard, publicly available datasets for agriculture are scarce, and representative data is important for machine-learning performance.

    Source: §1 · p. 1, §1.1 · p. 2

  • Method

    The authors manually captured mango-leaf images, then cleaned and resized them to create a ready-to-use dataset and benchmarked three deep-learning models.

    Source: §3.1 · p. 9, §5 · p. 10

  • Results

    The dataset contains 4,000 images of about 1,800 distinct leaves and covers seven mango diseases.

    Source: §1.2 · p. 3

  • Takeaways & Limitations

    The authors state that the dataset may support transfer-learning use for predicting mango diseases in other countries.

    Source: §5 · p. 10

    Dataset preparation faces challenges from disease diversity and from selecting useful validation techniques.

    Source: §2.3 · p. 8

Abstract

from arXiv · show

Agriculture is of one of the few remaining sectors that is yet to receive proper attention from the machine learning community. The importance of datasets in the machine learning discipline cannot be overemphasized. The lack of standard and publicly available datasets related to agriculture impedes practitioners of this discipline to harness the full benefit of these powerful computational predictive tools and techniques. To improve this scenario, we develop, to the best of our knowledge, the first-ever standard, ready-to-use, and publicly available dataset of mango leaves. The images are collected from four mango orchards of Bangladesh, one of the top mango-growing countries of the world. The dataset contains 4000 images of about 1800 distinct leaves covering seven diseases. Although the dataset is developed using mango leaves of Bangladesh only, since we deal with diseases that are common across many countries, this dataset is likely to be applicable to identify mango diseases in other countries as well, thereby boosting mango yield. This dataset is expected to draw wide attention from machine learning researchers and practitioners in the field of automated agriculture.

1. Introduction

The introduction identifies a shortage of representative agricultural datasets as a barrier to machine learning and presents a public mango-leaf dataset from Bangladesh as a response.

1. Introduction

  • Food production needs to rise by 70% by 2050, while plant diseases remain a barrier and machine learning offers a route to assist diagnosis.The introduction emphasizes that machine-learning performance depends on having an appropriately representative dataset.Source: §1 · pp. 1–2

1.1. Motivation

  • Existing plant-disease datasets are scarce, and the authors found no standard, sizable, publicly available dataset for mango-leaf diseases.They describe PlantVillage and DigiPathos and report finding no dataset of Bangladesh mango leaves.Source: §1.1 · p. 2

1.2. Contribution

  • The authors developed the first dataset of Bangladesh mango leaves, manually capturing and expert-labeling images before expanding the collection to 4,000 images.The collection began with 1,800 manually captured images; zooming and rotation brought it to 4,000.Source: §1.2 · p. 3
  • The dataset covers seven major mango diseases and is released publicly as a ready-to-use resource.The authors also report applying data-validation techniques to produce a processed dataset.Source: §1.2 · p. 3
  • The paper reports benchmark predictions from three machine-learning models fitted on the dataset.The authors also propose transfer learning as a way to use the Bangladesh-only dataset to predict mango diseases in other countries.Source: §1.2 · p. 3

2. Methodology

The methodology develops a mango-leaf dataset through disease study, orchard selection, image capture, validation, and preprocessing. It also examines disease traits in the images and practical challenges in dataset preparation.

2. Methodology

  • The dataset-preparation process includes studying prevalent diseases, consulting agricultural experts when selecting orchards, capturing leaves, and validating the images.The paper frames data quality and quantity as important to machine-learning performance.Source: §2 · p. 3, §2.1 · p. 3

2.1. Steps of Dataset Collection and Preparation

  • The preparation procedure moves from disease study and orchard selection to leaf capture and validation, including manual expert labeling, resizing, and background-noise cleaning.The authors also apply benchmark models to assess how they use the dataset.Source: §2.1 · p. 3

2.1.1. Studying the Common Diseases of Mango Leaves

  • The study focuses on seven mango diseases, including Anthracnose, Bacterial Canker, Cutting Weevil, Die Back, Gall Midge, Powdery Mildew, and Sooty Mould.Healthy leaves form an eighth category in the dataset.Source: §2.1.1 · p. 4, §2.1.3 · p. 4

2.1.2. Selecting the Mango Orchards for Data Collection

  • Four orchards in Dhaka, Savar, Rajbari, and Nilphamari were selected based on orchard size and tree variety, and yielded diseased leaves.Source: §2.1.2 · p. 4

2.1.3. Capturing the Leaf Images

  • Researchers located affected trees, sometimes with agricultural experts’ help, then photographed diseased and healthy leaves individually against a white background.Some diseases, including Bacterial Canker and Gall Midge, were relatively uncommon in the collected leaves.Source: §2.1.3 · p. 4

2.1.4. Validating the Dataset

  • 4,000 images across eight categories, with exactly 500 images per category, result from retaining about 1,800 of 6,000 captures and augmenting selected images.Images were resized to 240×320, manually cleaned, and some leaves with multiple disease traits were omitted.Source: §2.1.4 · p. 4

2.2. Traits of Different Diseases in the Leaf Images

  • Disease categories have distinguishable visual traits, such as Anthracnose’s black necrotic patches, Cutting Weevil’s cut-like leaves, and Gall Midge’s pimple-like growths.The authors describe these traits as making the images suitable for applying machine-learning models.Source: §2.2 · pp. 6, 8

2.3. Challenges to Prepare a Leaf Image Dataset

  • Preparation challenges include covering diverse diseases, choosing orchards and trees across a broad geographic spread, reaching leaves on tall trees, and selecting suitable validation techniques.The authors note that choosing useful validation methods requires data-science and machine-learning background.Source: §2.3 · p. 8

3. Applied Machine Learning Models and Result Analysis

The section benchmarks three machine-learning models on the dataset and reports CNN-SVM as the top performer.

3.1. Machine Learning Models

  • The benchmark uses CNN, ResNet50, and a CNN-SVM blend.Technical details and parameter settings are given in Appendix A.Source: §3.1 · p. 9

3.2. Result Analysis

  • The models are evaluated using precision, recall, and F1 measures.The section explains accuracy, precision, recall, and F1, with F1 combining precision and recall.Source: §3.2 · p. 9
  • CNN-SVM ranks above CNN and ResNet50 in the reported performance comparison.The supplied passage gives no metric values for the comparison.Source: Table 3 · p. 9, §3.2 · p. 9

4. Related Work

Prior work applies machine learning to plant-disease images, but the cited mango-leaf studies use limited or unavailable datasets. The authors identify a lack of standard, sizable, publicly available mango-leaf disease data.

4. Related Work

  • The cited mango-leaf studies use small proprietary datasets or cover only a limited set of diseases.Examples include 20 images for four diseases and fewer than 100 images for two diseases plus a healthy class; the latter dataset is not shared.Source: §4 · p. 9
  • The authors identify no prior investigation of mango-leaf diseases using a standard, sizable, publicly available dataset.They present their research as addressing this gap.Source: §4 · p. 10

5. Conclusion

The paper concludes by presenting a ready-to-use mango-leaf dataset and benchmark results as contributions to automated agriculture.

5. Conclusion

  • The dataset contains 4,000 images of about 1,800 distinct leaves, captured in four Bangladeshi orchards and covering seven diseases.The images were cleaned of background noise and resized; the authors also provide benchmark model performance and note the dataset could be enlarged with augmentation.Source: §5 · p. 10

A. Appendix: Technical Details of Applied Machine Learning Models

The appendix introduces technical descriptions and parameter settings for the three machine-learning models used in the study.

A. Appendix: Technical Details of Applied Machine Learning Models

A.1. Model 1: CNN

The first model is a CNN with five convolutional layers, three max-pooling layers, and a fully connected output layer. Its convolutional and pooling layers extract features while reducing feature-map resolution.

A.1. Model 1: CNN

  • Convolutional layers extract features from lines, edges, and corners at earlier layers and higher-level features at later layers.Source: §A.1 · p. 10
  • Pooling layers reduce feature resolution and overfitting tendency, while the fully connected layer produces the image’s final prediction.Source: §A.1 · p. 10
  • CNN neurons connect to a small portion of layers rather than all neurons, reducing the total number of neurons.Source: §A.1 · p. 10
  • The model uses five convolutional layers, ReLU after each, three max-pooling layers, and a final fully connected layer with 512 hidden neurons.Source: §A.1 · p. 10

A.2. Model 2: Resnet50

The second model is ResNet50, implemented with TensorFlow Keras default settings. Its shortcut paths support gradient flow, and the architecture comprises convolutional and pooling layers organized into five parts.

A.2. Model 2: Resnet50

  • ResNet50 has 48 convolution layers, one maximum-pool layer, and one average-pool layer, organized into five parts with convolution and identity blocks.Source: §A.2 · p. 10
  • ResNet50 trains about 23 million parameters and uses TensorFlow Keras with default parameter settings.Source: §A.2 · p. 10
  • A shortcut path lets gradients flow and is described as solving vanishing gradients, a problem attributed to changing backpropagation error signals.Source: §A.2 · p. 10

A.3. Model 3: Blend of CNN and SVM

The third model combines a CNN feature extractor with an SVM output layer. The SVM classifies the extracted feature set by finding a margin-maximizing hyperplane between two classes.

A.3. Model 3: Blend of CNN and SVM

  • The SVM transfers its input vector to a higher-dimensional feature space and finds a two-class hyperplane by maximizing the margin.Source: §A.3 · p. 11
  • The authors motivate the hybrid by noting that a traditional machine-learning model as a CNN’s last layer is often found to yield better performance.Source: §A.3 · p. 11
  • The CNN-SVM model uses five convolutional layers, three 2×2 max-pooling layers, a 512-neuron fully connected layer, and an SVM output layer.Source: §A.3 · p. 11
  • Flattened CNN features form each image’s feature set and are fed to the fully connected layer before SVM classification.Source: §A.3 · p. 11

A.4. Implementation Details

The three models were implemented in TensorFlow using Python on Google Colab with GPU execution. Training used model-specific losses, Adam, a batch size of 128, and 100 epochs, with image augmentation during training.

A.4. Implementation Details

  • The models were implemented in Python with TensorFlow on Google Colab and executed on a GPU.Source: §A.4 · p. 11
  • All three models used a batch size of 128 and trained for 100 epochs with Adam; CNN and ResNet50 used categorical cross-entropy, while CNN-SVM used squared hinge loss.Source: §A.4 · p. 11
  • TensorFlow image data generation applied scaling, shearing, zooming, and flipping, so training used more than the dataset’s 4,000 original images.Source: §A.4 · p. 11
Loading 2209.02377v1…