An automated plant leaf disease classification framework using three-step feature extraction-based residual Bi-RNN with spatial channel attention

The demand for crop production is highly increased because of distinct disorders that endanger food security. Hence, accurately recognizing plant disorders is crucial and significant in the agricultural sector. The conventional classification mechanisms such as laboratory examinations, and naked-eye observation, have distinct limitations including subjective and time-consuming. Nowadays, deep learning strategies have attained widespread application in the classification of plant diseases. These models have resolved the issues of the conventional model’s limitations. However, the existing automated models also have some primary drawbacks such as high computational burdens, expenses, and inaccurate outcomes. Considering these challenges, a new computer-aided system for rapid detection and diagnosis system is introduced. At first, the plant leaf images are fetched from the data sources and these images are given to the image pre-processing. Further, the pre-processed images are subjected to the leaf segmentation phase, where the optimal binary thresholding is supported. Here, the parameter optimization is done through the Self Adaptive Pelican Optimization Algorithm (SA-POA). After segmentation, a three-step feature extraction is conducted. Here, the Convolutional Neural Network (CNN) technique is supported for extracting the first set of features. Further, the second feature set such as Local Gradient Patterns (LGP) and Local Binary Patterns (LBP) are extracted. Finally, the third set of features such as the Gray Level Cooccurrence Matrix (GLCM) and shape features are fused for improving the classifier function. Further, fused feaure is subjected to the Residual Bi-directional Recurrent Neural Network with Spatial Channel Attention (RBRNN-SCA) technique for classifying plant leaf diseases. The experiments are validated using diverse performance measures to show the effective outcomes. Moreover, the suggested classification model has attained 95% accuracy than the existing classifiers. Throughout the entire validation, the suggested method outperforms superior performance.