Development of a Mobile Solution for Offline Disease Diagnosis in Moringa Leaves Using Convolutional Neural Networks

Abstract

Mobile-based disease detection has become an important tool for crop health monitoring in resource-limited farming settings. Moringa oleifera is a fast-growing, drought-resistant, and highly nutritious tree that is nonetheless vulnerable to leaf diseases such as leaf spot, bacterial blight, powdery mildew, insect damage, rust and nutritional deficiency, yet very few diagnostic tools exist for this underutilised crop. This study presents the development of an offline mobile expert system for Moringa leaf disease diagnosis that combines convolutional neural networks (CNNs) with cross-platform mobile deployment. A total of 110 Moringa leaf images were collected, of which 102 images spanning six disease categories and a healthy class were pre-processed in MATLAB using resizing, Contrast Limited Adaptive Histogram Equalisation, Gaussian blur and RGB normalisation, then used to train and compare six pretrained CNN architectures—SqueezeNet, MobileNetV2, ResNet-18, ResNet-50, ResNet-101 and DenseNet-201—against accuracy, training time and model size. SqueezeNet reached the highest hold-out validation accuracy (52.63%, shared with MobileNetV2 and ResNet-101) with the shortest training time and the smallest model, around 95% smaller than ResNet-50, and was therefore selected for deployment. When the trained models were applied to all 102 dataset images, which include the training images, SqueezeNet, MobileNetV2 and ResNet-18 achieved 89.22%, 82.35% and 90.20%, respectively. Retaining RGB colour information improved accuracy by more than 10% relative to greyscale input. SqueezeNet was converted to ONNX and then to TensorFlow Lite, and embedded in a Flutter-based Android application designed for fully offline inference together with a local SQLite disease-and-treatment knowledge base. Mobile-integration testing confirmed offline image selection and on-device processing but also revealed implementation challenges, including intermittent inference failures linked to image-tensor handling in the Flutter runtime. The results demonstrate that lightweight CNNs can be practically deployed for offline plant disease diagnosis and support the broader goal of delivering accessible, data-driven agricultural tools to smallholder and rural Moringa farmers