Dual RNN Architecture for Crop Disease Detection: Improved Patch-Based CNN for Segmentation

Abstract

Crop diseases pose a significant threat to global food security, leading to substantial yield losses and economic repercussions. Timely intervention and efficient management of many disorders depend on their prompt and precise detection. Deep learning techniques have transformed computer vision in recent years and provided potential solutions for automated crop disease diagnosis. The proposed framework of dual RNN-based crop disease detection (D-RNN-based CDD) encompasses several crucial stages aimed at harnessing deep learning methods for precise and efficient disease identification in agricultural crops. First, preparation methods are used to improve the raw images quality. This involves the contrast transformation to enhance visibility and the application of a Gaussian filter to reduce noise, ensuring smoother images suitable for subsequent processing. After preprocessing, image segmentation is done using a hyper softmax patch-based convolutional neural network (HSP-CNN) approach, dividing the images into smaller patches for localised analysis. This segmentation method allows the model to focus on specific regions of interest, facilitating accurate identification of diseased areas. Following segmentation, the next step is feature extraction that captures pertinent characteristics indicative of crop diseases. Shape features, such as area, convexity, centroid and perimeter, are extracted to quantify the disease symptoms. Colour features capture RGB components in terms of mean, median and standard deviation. Moreover, texture features are extracted to analyse the patterns and structures associated with different diseases using modified texture orientation-based multitexon (MTOM) features enabling extraction at multiple scales. These extracted features offer comprehensive representations of the underlying characteristics of diseased regions within crop images. Finally, classification is done using a dual RNN (D-RNN) comprising two triple hidden layers assisted in RNN (THA-RNN) models. One THA-RNN model trains the segmented images, while the other captures the extracted feature set and trains on it. Leveraging the capabilities of THA-RNNs, adept at analysing sequential information (features), the classifier accurately predicts the presence or absence of crop diseases. For Dataset 1, the developed THA-RNN model achieved the maximum value of 0.955, outperforming other models, such as GoogLeNet (0.911), LeNet (0.921), CNN (0.894), Bi-LSTM (0.877), DNN (0.895), RNN (0.873), RESNET (0.886) and ANN (0.870).