Automated systems for detecting plant diseases play a pivotal role in precision agriculture, crucial for maintaining crop health and maximizing yields. This study specifically targets the early identification and severity classification of coffee leaf diseases, focusing on the Arabica variety. Leveraging the superior image classification capabilities of ResNet-50, the methodology introduces a system designed to differentiate between healthy leaves and those experiencing biotic stress such as rust, miner Cercospora, and Phoma. In cases of disease presence, the system further categorizes the severity level. The process commences with the collection and preprocessing of a diverse dataset containing coffee leaf images depicting various health states. Employing data augmentation and visualization techniques, such as the Grad-Cam method, enhances the model’s robustness and accurately highlights regions within coffee leaf images for precise classification decisions. A modified ResNet-50 model is utilized for classifying leaf health status and identifying biotic stress, followed by severity grading employing the Grad-Cam technique.