The surge in agricultural output is inevitable due to the global population growth, which threatens food security by 2050. As a key staple, tomatoes are vulnerable to diseases that adversely affect their yield and quality. Therefore, effective disease management is vital for the economic viability of tomato farming. This study proposes a new approach using a multi-scale ensemble-attention (MSEA) mechanism to detect tomato leaf diseases (TLD) accurately. The unique aspect of MSEA is the use of a hierarchical feature fusion strategy with an attention-based weighting mechanism that blends multi-scale features from the modified InceptionResNetV2 architecture equipped with an added convolutional layer for feature refinement. MSEA uses the modified InceptionResNetV2 architecture for multi-scale feature extraction to recognize different dis ease patterns, improving recognition results. Using element-wise multi plication followed by a convolutional layer, MSEA integrates channel and spatial attention. This allows the model to focus on disease symptoms. Grad-CAM visualization improves interpretability. MSEA demonstrates an accuracy of 98.87%, a specificity of 99.95%, and an MCC of 0.9902, surpassing DVTXAI’s 93.56% accuracy. Although MSEA’s accuracy is comparable to XSE-TomatoNet’s 98.88%, it shows superior specificity and MCC in adverse conditions like fluctuating lighting. These results demonstrate that MSEA offers a promising solution for improved crop yields and sustainable agriculture.