PaperlayerView exact revision on arXiv

Source-linked AI summary

Leaf Image-based Plant Disease Identification using Color and Texture Features

Nisar Ahmed, Hafiz Muhammad Shahzad Asif, Gulshan Saleem

arXiv:2102.04515v1cs.CV

TL;DR

Delayed plant-disease identification can reduce yield, while deep-learning approaches may be too computationally demanding for handheld devices. This paper proposes a two-stage, feature-based leaf-image classifier and reports 99.31 accuracy and 99.33 F1-score.

  • Problem

    Delayed plant-disease identification can reduce yield, while deep-learning approaches require more computation than resource-limited handheld devices support.

    Source: §I · p. 1

  • Method

    A two-stage system classifies leaves as healthy or diseased, then segments diseased regions and uses selected color and texture features for disease classification.

    Source: §II · p. 4, C. Segmentation of Diseased Region · p. 8

  • Results

    The proposed model reports 99.31 accuracy and 99.33 F1-score.

    Source: §IV · p. 20

  • Takeaways & Limitations

    The approach provides results comparable to state-of-the-art algorithms and demonstrates the effectiveness of texture-based visual classification for disease identification.

    Source: §II · p. 4

Abstract

from arXiv · show

Identification of plant disease is usually done through visual inspection or during laboratory examination which causes delays resulting in yield loss by the time identification is complete. On the other hand, complex deep learning models perform the task with reasonable performance but due to their large size and high computational requirements, they are not suited to mobile and handheld devices. Our proposed approach contributes automated identification of plant diseases which follows a sequence of steps involving pre-processing, segmentation of diseased leaf area, calculation of features based on the Gray-Level Co-occurrence Matrix (GLCM), feature selection and classification. In this study, six color features and twenty-two texture features have been calculated. Support vector machines is used to perform one-vs-one classification of plant disease. The proposed model of disease identification provides an accuracy of 98.79% with a standard deviation of 0.57 on 10-fold cross-validation. The accuracy on a self-collected dataset is 82.47% for disease identification and 91.40% for healthy and diseased classification. The reported performance measures are better or comparable to the existing approaches and highest among the feature-based methods, presenting it as the most suitable method to automated leaf-based plant disease identification. This prototype system can be extended by adding more disease categories or targeting specific crop or disease categories.

I. INTRODUCTION

Delays in plant-disease identification can reduce yield, while prior approaches differ in accuracy, computational demands, and coverage. This work proposes visual leaf-image classification using a staged, feature-based approach across PlantVillage’s 38 classes.

I. INTRODUCTION

  • Delays in plant-disease identification reduce yield quantity and quality, and earlier studies often covered few crops or diseases with small samples.The introduction identifies timely identification as important for maximum yield.Source: §I · pp. 1–2

II. LITERATURE REVIEW

  • Feature-based methods use hand-crafted features, while deep-learning approaches report higher accuracy but require more computation and are unsuitable for resource-limited handheld devices.The cited literature also describes prior systems targeting particular plants or diseases.Source: §II · p. 3
  • The proposed two-stage system first classifies leaves as healthy or diseased, then segments diseased regions and classifies disease using selected color and texture features with a cubic-kernel SVM.The authors report comparable results to state-of-the-art algorithms and emphasize texture-based visual classification.Source: §II · p. 4

III. MATERIALS AND METHODS

  • PlantVillage contains 54,309 leaf images across 38 classes and 14 crops; the system categorizes images only when disease is visually distinguishable.The dataset includes 26 disease classes and 12 healthy plant categories.Source: §III · pp. 4–5

A. Background Removal

The pipeline removes background before extracting leaf information, first separating healthy from diseased leaves and then segmenting diseased regions for disease identification. Hue-based watershed processing supports background removal, while Otsu thresholding isolates diseased areas; segmentation is imperfect in some images.

A. Background Removal

  • Background removal transforms RGB images to HSV and uses the Hue layer, followed by bilateral filtering, watershed segmentation, and morphological processing.The authors state that Hue excludes intensity-related information, while RGB channels and true-color images contain shadows and reflections that hinder background removal.Source: A. Background Removal · pp. 6–7, Figure 2 · p. 6
  • Most images are segmented well, but some contain extra nonleaf area and the squash powdery-mildew class has particularly poor background removal.The authors report that slightly poor segmentation does not drastically affect disease classification.Source: A. Background Removal · p. 7

B. Healthy and Diseased Classification

  • Disease identification first classifies leaves as healthy or diseased, then categorizes diseased leaves among 26 disease classes.The healthy-versus-diseased step uses SURF features and a bag-of-features SVM; subsequent processing is applied only to diseased leaves.Source: B. Healthy and Diseased Classification · p. 8

C. Segmentation of Diseased Region

  • Otsu’s algorithm segments diseased regions to extract discriminative texture features, using a computationally inexpensive approach.The resulting segmentation is described as satisfactory, though not exceptionally accurate; Figure 6 shows examples for corn, grape, and potato.Source: C. Segmentation of Diseased Region · pp. 8–9, Figure 6 · p. 9

D. Feature Extraction

The method represents diseased leaf regions with compact RGB color moments and texture descriptors derived from a grayscale GLCM. Texture is treated as the more informative signal for disease identification.

D. Feature Extraction

  • The feature set emphasizes color and texture because these vary across disease classes, while shape is considered less useful for disease identification.The authors describe color as similar within classes and different between classes, and texture as dependent on disease type.Source: D. Feature Extraction · p. 9

Color Features

  • Six color features summarize the diseased region: RGB means and standard deviations, with color moments used as a compact representation.The paper describes color as a lesser predictor of disease categories than texture.Source: D. Feature Extraction · pp. 9–10

Texture Features

  • Texture is considered more informative for disease identification, with 22 features extracted from an eight-level grayscale GLCM of the diseased region.The 8 × 8 matrix records how frequently pixels at gray levels i and j occur adjacent to one another.Source: D. Feature Extraction · pp. 10, 12

E. Feature Scaling

Feature scaling addresses the different ranges of input features, which can disproportionately affect distance-based calculations. The method standardizes features to zero mean and unit variance.

E. Feature Scaling

  • Standardization scales features to zero mean and unit variance, the method selected here for SVM, ANN, and linear regression.Feature scaling helps prevent wide-ranging features from disproportionately affecting distance measures; the equation uses the feature value, mean, and standard deviation.Source: E. Feature Scaling · p. 12

F. Feature Selection

Feature selection compares forward sequential selection with ReliefF, while classifier selection evaluates five algorithms and tunes the best performer. Forward feature selection (FFS) produces the better-suited feature subset for classifier training.

F. Feature Selection

  • Feature selection addresses correlated color and texture features, since including all features may degrade performance and slow training.The study explored forward feature selection and ReliefF.Source: F. Feature Selection · p. 13
  • FFS selected 19 features and reached a 10-fold cross-validation accuracy of 0.934; its subset was better suited to classifier training than ReliefF’s eight-feature subset.The two subsets were compared for classification performance.Source: F. Feature Selection · pp. 13–14

G. Classifier Selection

  • Five classifiers were assessed—multi-SVM, K-NN, Naïve Bayes, Random Forest, and ANN—and the best performer was optimized for its hyperparameters.The study used one-vs-one multi-SVM and found the cubic kernel had the best cross-validation performance among the tested kernels.Source: G. Classifier Selection · pp. 14–15
  • Random Forest averages predictions from trees trained on replacement-sampled subsets, addressing decision-tree overfitting, but has slower prediction speed.The passage also characterizes the algorithm as complex.Source: G. Classifier Selection · p. 15

IV. RESULTS AND DISCUSSIONS

The final disease-classification model reports 99.31 accuracy and 99.33 F1-score, while the aggregate two-stage results are slightly lower. Evaluation also compares feature subsets and classifiers and tests the model on independently collected images.

IV. RESULTS AND DISCUSSIONS

  • Forward feature selection produced slightly better accuracy than the complete feature set, whereas the ReliefF subset performed significantly worse.The authors report that forward feature selection also improved performance in time and classification accuracy.Source: §IV · pp. 16–17
  • The ANN configuration search yielded 93.4% classification accuracy for the network with the least cross-entropy.Configurations varied the neuron counts in one- and two-hidden-layer networks.Source: §IV · p. 17, Table 8 · p. 17, Figure 7 · p. 17
  • The cubic-kernel multi-SVM had higher accuracy than the tested ANN with 50 neurons in each of two hidden layers.The cubic-kernel SVM was selected as the final model, prioritizing prediction speed for practical use because training is a one-time job.Source: §IV · pp. 17–18, Table 9 · p. 18
  • 99.31 accuracy and 99.33 F1-score are reported for the proposed disease-classification model across 26 diseased classes.The full dataset contains 38 classes; the 26-class results are not directly comparable with studies classifying only leaf categories.Source: §IV · p. 20
  • The proposed approach uses two classification stages, and the final aggregate score is slightly below the disease-classification module’s score.The first stage distinguishes healthy from diseased leaves; the second classifies diseased leaves into disease categories.Source: §IV · pp. 20–21
  • The ANN’s accuracy changed slightly across repeated training, and the search was limited by computational complexity.The authors associate the variation with weight initialization and difficulty finding a global minimum.Source: §IV · p. 20
  • The proposed two-stage results are not directly comparable with other methods, and healthy-leaf category classification is outside the study’s scope.The reported cross-validation comparison used 10 train-test splits; the disease-category stage covers diseased leaves only.Source: §IV · pp. 20–21
  • 82.47% disease-category accuracy and 91.40% healthy-and-diseased accuracy were achieved on the self-collected dataset.The authors attribute the lower disease-category accuracy to difficulty segmenting diseased leaf areas.Source: §IV · p. 21

V. CONCLUSION

The model first classifies leaves as healthy or diseased, then identifies disease in unhealthy leaves using selected color and texture features. It reports 98.79% disease-classification accuracy in 10-fold cross-validation and evaluates generalization on a self-collected dataset.

  • 98.79% classification accuracy, with a standard deviation of 0.57, was achieved for disease classification over 10-fold cross-validation.Source: §V · p. 22
  • The two-stage model first classifies leaves as healthy or diseased, then identifies the disease affecting unhealthy leaves.The first stage uses bag-of-features; the second extracts diseased leaf areas and uses selected color and texture features.Source: §V · p. 22
  • On a self-collected dataset, accuracy was 91.40% for healthy-versus-diseased identification and 82.47% for disease-category identification.Source: §V · p. 22
  • The prototype classifies only visually distinguishable images represented in its training database.Source: §V · p. 22

A. Future Work

Future work proposes integrating plant-species identification before disease identification and developing specialized crop models that also categorize disease severity.

A. Future Work

  • An integrated system could identify plant species from the leaf image in its first stage and plant disease in its second.Source: A. Future Work · p. 22
  • Specialized models for high-value or other specific crops could identify diseases and categorize their severity through fine-grained recognition.Source: A. Future Work · p. 22

APPENDIX A

Appendix A describes texture features that characterize gray-level distributions and relationships in segmented diseased leaf regions.

APPENDIX A

  • Uniformity measures the orderliness of gray-levels, while entropy measures image disorder and information content.Source: Appendix A · p. 24
  • Contrast represents gray-level variation, while homogeneity measures smoothness and is usually larger when contrast is smaller.Source: Appendix A · pp. 24–25
  • Inverse Difference is another measure of smoothness; correlation measures neighboring gray-level dependency, and autocorrelation measures correlation with a shifted region.The shift amount is denoted by k.Source: Appendix A · p. 25
  • Cluster shade uses a third-order moment and has a compact range for homogeneous regions; cluster prominence uses a fourth-order moment with similar characteristics.Source: Appendix A · p. 26
  • Sum Average measures the mean of the gray-level sum distribution, while Sum Variance measures its dispersion around the mean.Source: Appendix A · p. 26
  • Sum Entropy measures disorder in the gray-level sum distribution; Difference Variance measures dispersion in the gray-level difference distribution.Source: Appendix A · p. 27
  • Difference Entropy measures disorder in the gray-level difference distribution, and information measures of correlation are described as rectangular correlation measures.Source: Appendix A · p. 27
  • Maximal Correlation Coefficient measures dependence between two random variables; IDN and IDMN are normalized homogeneity measures, with IDN normalized by the eight gray-levels used.IDN and IDMN are described as closely related to contrast and dissimilarity.Source: Appendix A · pp. 27–28

APPENDIX B

Appendix B defines classifier evaluation measures and explains how to read a confusion matrix.

APPENDIX B

  • A confusion matrix compares actual classes in columns with predicted classes in rows, revealing classes the classifier may confuse.Source: Evaluation Measure · p. 28
  • Recall is the share of relevant instances that are retrieved, while F1-score is the harmonic mean of precision and recall.F1-score is noted as useful when class distribution is uneven because it accounts for false positives and false negatives.Source: Evaluation Measure · pp. 28–29
  • Accuracy is the ratio of correctly classified observations to total observations; precision is the share of retrieved examples that are relevant.Source: Evaluation Measure · p. 28
Loading 2102.04515v1…