Benchmarking CNN Models for Black Pepper Diseases and Malnutrition Prediction

Black Pepper (Piper nigrum L.) is one of the most important commodities in Southeast Asia. Like other plants, black pepper plants are exposed to various diseases and growth issues. Automated recognition of plant diseases and nutrient deficiencies is important with the availability of mobile devices. This paper aims to provide a benchmark for CNN models in learning the symptoms related to black pepper disease or malnutrition indicated on the leaves. Samples of black pepper leaf for 11 diseases and nutrient deficiencies were collected from farms in Sarawak. A total of 1,043 images of the samples were taken in a controlled environment. Augmentation was performed to increase the number of samples and to generate variation. Five deep-learning neural network architectures were selected for the modelling of the classification task. The results showed that state-of-the-art CNNs EfficientNet-B0 (0.88 accuracies), MobileNet-v2 (0.88 accuracies), ResNet-V2-50(0.86 accuracies), DenseNet121 (0.85 accuracies) and a customized CNN (0.85 accuracies), can act as benchmarks for black pepper disease or malnutrition diagnosis via the leaves. The models have the potential to be used for early detection and intervention for black pepper plant management.