Plant diseases segmentation is critical for fine-grained detection and precision agriculture. However, the opacity of the decision-making process limits trust between Artificial Intelligence (AI) and humans. This study proposes the eXplainable AI for Multi-class Plant Leaf Disease Segmentation (XAI-MPLDS) approach using a fine-tuned transformer-based architecture (SegFormer) on a custom dataset. To the best of our knowledge, this is the first approach integrating XAI into semantic segmentation for plant disease detection. To add interpretability, four complementary post-hoc explanation methods —Grad-CAM, Grad-CAM++, LIME, and RISE—were customized. Each method provides insights into the model’s decision-making process. The experimental results indicate that the gradient-based showed sharp localization of class activations consistent with the model’s attention mechanism, while the perturbation-based offered complementary perspectives on feature importance. As for the quantitative evaluation, RISE achieved the highest fidelity scores in terms of IoU and Dice as 82% and 89%, followed by Grad-CAM++ as 78% and 81%, Grad-CAM as 73% and 82%, and LIME as 32% and 34% respectively. These findings demonstrate the value of customizing multiple XAI techniques to gain a multi-perspective comprehensive understanding of the semantic segmentation model’s behavior in complex, real-world segmentation tasks, where farmers and experts can trust and understand AI-driven solutions supporting early and precise localization of plant pathology, precision treatment and plant phenotyping.