<p>Crops are a major food source for the world; however, their productivity is severely threatened by pests and diseases, resulting in significant yield losses and economic damage. Visual inspection by farmers often delays diagnosis, allowing pests and diseases to rise to more advanced and severe stages. To address such challenges, and enable timely, generalizable diagnosis, this paper proposes a SmoothGrad-Integrated Gradients explainable EfficientRefineNet architecture. The EfficientRefineNet augments the EfficientNet-B0 baseline by embedding convolutional blocks, which are ameliorated with a sequential channel-first and spatial attention mechanism, enabling sequential refinement of feature maps. This design significantly strengthens the capacity of the network to localize and emphasize the most informative features, thereby enhancing the effectiveness of crop pest and disease diagnosis. Then, SmoothGrad-Integrated Gradients (SGIG) explainable Artificial Intelligence (AI) method is integrated to improve interpretability by highlighting the image areas that most influence EfficientRefineNet’s diagnosis of crop pests and diseases. The proposed architecture is trained and validated on the CCMTO dataset, which is constructed by combining the Cashew, Cassava, Maize, and Tomato (CCMT) dataset with the Onion crop samples from the TOM2024 dataset. Its performance is evaluated on three testing sets: CCMTO, TOM2024 excluding Onion, and PlantVillage, to verify robustness across diverse crop types and conditions. Cloud-based evaluation is performed by testing the trained model on Google Colab and deploying it as an interactive application on the Hugging Face platform, demonstrating its usability and practical potential for timely and widely accessible diagnosis. Experimental results using multiple evaluation metrics confirm that EfficientRefineNet significantly outperforms current state-of-the-art architectures.</p>