Hybrid deep learning system for crop disease classification using modified SegNet segmentation

Abstract

In traditional agricultural systems, managing crop diseases faces significant challenges, primarily due to the reliance on visual inspection and manual symptom identification. These methods are often time-consuming, error-prone, and may fail to detect diseases early or accurately, leading to ineffective treatments and substantial crop loss. Furthermore, the unpredictability of disease symptoms and the existence of similar-looking diseases complicate diagnosis. To address these limitations, there is a growing necessity for innovative deep learning-based methods. This study proposes an advanced Modified LinkNet-Bidirectional Long Short-Term Memory (MLBLSTM)-based system for crop disease classification, incorporating a multi-step process starting with data collection from three datasets: apple, corn, and pepper plant leaves. The preprocessing phase utilizes Enhanced Wiener Filtering (EWF) to preserve high-frequency details and enhance image quality. The filtered images are processed through an advanced Modified SegNet (MSegNet) model to do the segmentation process. Feature extraction follows, leveraging Hierarchy of Skeleton (HOS), Modified Local Gabor Increasing Pattern (MLGIP), Median Binary Pattern (MBP), and statistical features. Finally, the classification step employs a hybrid model combining Modified LinkNet (MLNet) with a novel σ-SE block and Bidirectional Long Short-Term Memory (Bi-LSTM) classifiers. The validation results prove the performance of MLBLSTM model measures with an accuracy of 0.947, a sensitivity of 0.955, and a specificity of 0.936.