Classifying Nutritional Deficiencies in Coffee Leaf Using Transfer Learning and Gradient-Weighted Class Activation Mapping (Grad-CAM) Visualization

Coffee is a very popular beverage around the world. The production coffee can be affected by nutrition deficiencies and result in massive underproduction. This makes early detection of nutrition deficiency so important. In this research we will use transfer learning approach to create an automated system that can detect and classify different kind of nutrition deficiency from coffee leaf images. The dataset used in the research has 1006 leaf images with 10 classes in total; one of which is a healthy class, The dataset was pre-processed thoroughly by applying proper techniques. DenseNet 201, InceptionV3 and Xception, these three CNN models have been utilized for the classification task. The deep learning models were fine-tuned by freezing model layers, and hyperparameters were tuned. Those transfer learning models both mainly work as feature extractor with only their last convolutional block being trainable and a custom model of our own was used for classification. The Xception model achieved the highest accuracy of 83.32%. To interpret how the model is working we integrated Explainable AI (XAI) named Gradient-Weighted Class Activation Mapping (GRAD-CAM).