Toward Sustainable AI in Agriculture: An Efficient VQA Pipeline for Rice Leaf Disease Analysis

Visual Question Answering (VQA) has gained increasing attention as a multi-modal AI approach capable of interpreting visual content through natural-language queries. However, its application in agriculture remains underexplored, particularly in crop disease diagnosis, where timely and interpretable decision support is essential. To address this gap, we introduce a new VQA dataset specifically developed for rice leaf disease identification, covering three major disease classes and consisting of 7,570 carefully curated question–answer pairs. Comparative experiments reveal that the pipeline with distilled models, DeiT and DistilBERT, achieves an accuracy of 0.5064 and WUPS of 0.5056, matching the performance of ViT + BERT while reducing trainable parameters by 20% and lowering GPU memory usage by 10%. These results demonstrate that lightweight architectures can deliver competitive reasoning capabilities with substantially reduced resource demands. By minimizing energy consumption and enabling deployment on edge devices, the proposed pipeline aligns with the principles of sustainable AI, addressing the environmental and operational challenges of smart agriculture.