SKGDC: Effective Segmentation Based Deep Learning Methodology for Banana Leaf, Fruit, and Stem Disease Prediction

In agriculture, detecting plant diseases is crucial for optimal plant growth. Initially, input images are collected from three datasets: banana leaf spot diseases (BananaLSD) dataset, banana leaf dataset, and PSFD-Musa dataset. These collected images undergo filtering using the upgraded self-guided filtering (Up-SGF) model, chosen for its ability to preserve sharp edges effectively with strong smoothing, enhancing edge perception. Additionally, the improved Wiener convolution filter (IWCF) is applied to enhance image quality by reducing noise while preserving features. Following pre-processing, the images are segmented using spatial kernelized gravity-based density clustering (SKGDC) to identify disease-affected regions. Kaze, Blob, and histogram of oriented gradient (HOG) methods are utilized to extract relevant features: Kaze for scale-invariant key points, Blob for identifying and characterizing areas with shared color or intensity characteristics, and HOG for detecting edges and contours. These extracted features are then fused together using convolutional pyramid fusion (CPF). Finally, the hybrid channel attention-based YOLOv8 (HCA-YOLOv8) model is employed to predict diseases in banana fruit, stems, and leaves. Channel attention is incorporated into the backbone network architecture, and an improved deformable convolution replaces the normal convolution. Hyper parameters of the network model can be optimized using the improved chimp optimization (ICO) algorithm. Simulation results are obtained using the Python programming language. The proposed model attained 98.12% of accuracy in BananaLSD dataset, 99% of accuracy in banana leaf dataset, and 98.5% of accuracy in PSFD-Musa dataset.