<p>Cotton plants are susceptible to various diseases that significantly impact yield and quality. The current work extends the concept of a stacked convolutional autoencoder beyond the traditional sequential stacking to complementary feature-level stacking. The encoder component of the proposed autoencoder utilizes two distinct convolutional backbones- a residual-based encoder built on ResNet50 architecture and VGG-based encoder built on VGG 16 architecture, both integrated with Convolutional Block Attention Module (CBAM). Both the encoder components operate concurrently. Latent representations generated by both the encoders are stacked through feature level attention-based fusion. A lightweight fully connected classification component operates on the fused latent representation to map the compact latent features of cotton leaf to the disease classes. This design reduces computational complexity while ensuring that classification is performed on the most informative and discriminative features extracted by the autoencoder. Furthermore, to ensure interpretability, the framework incorporates Gradient-weighted Class Activation Mapping (Grad-CAM) – an explainable AI technique, is utilized to ensure the explainability and traceability of the decision-making process. Experimental results demonstrate that the proposed framework effectively leverages both hierarchical and complementary feature stacking, leading to accuracy varying from 97% to 99% for both the cotton plant disease datasets under study, compared to traditional stacked autoencoder approaches whose accuracy varies between 90% and 98%. Further the Grad-CAM visualizations demonstrate that the proposed dual-branch architecture learns complementary features, global structural abnormalities as well as fine-grained texture variations, associated with disease.</p>