Crop diseases are a menacing challenge to agricultural production, economic sustainability and ecological balance especially in crop-based economies. Detection methods currently employed are plagued by scalability and domain transferability problems which are introduced by dataset variability. In this work, a comprehensive artificial intelligence framework for multi-crop disease detection is introduced through the fusion of four heterogeneous datasets–PlantVillage, Sugarcane Plant Disease Dataset, Wheat Leaf Dataset, and Cotton Crop Disease Detection Dataset–into a large cross-domain benchmark. Utilizing an ensemble of five state-of-the-art transfer learning models (DenseNet169, EfficientNetB1, EfficientNetB2, EfficientNetV2B2, and EfficientNetV2B3) the proposed architecture successfully combines hard and soft voting mechanisms to improve prediction robustness, with an overall accuracy of 97%. This performance is superior to that of individual baseline models and current methods and provides better computational efficiency. For transparency of AI-driven decision-making, Local Interpretable Model-agnostic Explanations (LIME) are employed, providing interpretable explanations for predictions. The framework’s three main contributions are: the introduction of a new cross-crop dataset that is designed to facilitate accurate multi-class disease classification across multiple crops, an accuracy-optimized and efficiency-optimized ensemble strategy and the integration of explainable AI (XAI) to provide actionable diagnostic recommendations. Experimental results show unparalleled effectiveness in disease detection, with accuracy rates of 97% for early-stage infections, illustrating the framework’s utility in diverse agricultural settings.