AUXILIARY GENERATIVE ADVERSARIAL NETWORK (AGAN) BASED IMAGE GENERATION AND MULTI-SCALE ENCODER DECODER SELF-ATTENTION NETWORK (MSEDSAN) FOR DISEASE CLASSIFICATION IN COCONUT TREE

Abstract

Coconut trees are vital fasten crops that are widely grown in coastal regions, contributing significantly to the national economy. However, diverse types of diseases affect coconut tree health, including leaf spot diseases, which affect crop health and productivity. Deep learning algorithms that extract discriminative features from the images implicitly gained popularity in computer vision. In this paper, Auxiliary Generative Adversarial Network (AGAN) based image augmentation is employed to generate synthetic images, improving dataset diversity and reducing class imbalance. The augmented images are then processed using Transfer Learning-based Enhanced Visual Geometry Group 16 (EVGG16) to extract local features and a Swin Transformer to capture global contextual information. The extracted features are subsequently processed using the proposed Multi-Scale Encoder Decoder Self-Attention Network (MSEDSAN) to learn comprehensive and discriminative disease representations at multiple scales. MSEDSAN, Cross-Scale Attention (CSA) establishes interactions among shallow and deep feature representations to effectively integrate fine local disease features with high-level semantic information. Furthermore, a Spatial-Channel Attention Fusion (SCAF) adaptively emphasizes disease-relevant regions and informative feature channels while suppressing complex backgrounds and redundant information. The resulting features are further processed through a Multi-Scale Feature Enhancement (MSFE) to enhance disease-specific representations across multiple scales. Finally, Global Average Pooling (GAP) transforms the refined feature maps into a compact feature vector, which is passed through a fully connected layer with Softmax activation to classify the coconut leaf images into healthy and different disease categories. Proposed approach achieves improved disease classification performance and provides more reliable and explainable predictions for practical agricultural applications