Crop disease has significantly affected the production and commercialisation of watermelon fruits, causing a drop in food security, agricultural economy and rural development. Moreover, the traditional manual method of disease classification is time consuming and not completely efficient for large scale farming. A significant means to solving this crop disease wchallenges is the use of deep learning models like VGG16 architectures. Despite their high accuracy level, they are computationally demanding and not appropriate to address the challenge of watermelon disease classification within a resource constrained environment. Consequently, a comprehensive review of related literatures was conducted and the research gaps identified, led to the development of a lightweight model for watermelon diseases classification deployable to android mobile device to support real time decision-making on farm field. It utilises the transfer learning approach with MobileNetV2, a lightweight convolutional neural network (CNN) architecture renowned for striking a balance between high accuracy and computational requirements. The dataset comprises of high-quality images of different disease classes of watermelon namely, Anthracnose, Downy Mildew, Mosaic Virus and Healthy class. Exploratory data analysis (EDA) conducted on the dataset revealed the class distribution and imbalance which facilitated the use of augmentation to increase data diversity and generalisation.The model development was carried out following the agile framework and software development life cycle (SDLC) which divided the processes into structured iterative sprints ranging from requirement analysis to model deployment and testing on android device. The model training was achieved in two phases, Feature extraction of ImageNet – MobileNetV2 pretrained weights and fine tune of the last 100 layers of the CNN architecture. The effect of batch sizes (16, 32, and 64) on the model accuracy was examined during model development and batch size of 16 performed best in terms of efficiency and generalisation with a validation accuracy of 99%.The model deployment capacity was evaluated with a different test data curated from extension service website. The TensorFlow lite version of the developed model had a size of 9.08MB while the total inference speed on 27 images of Anthracnose class and training time was 0.23 and 10634.26 seconds respectively with an overall classification accuracy of 56% on model developing platform and 61% on developed mobile application built with Kotlin and Jetpack Compose. The test evaluation identified anthracnose disease to have a higher classification accuracy which is probably due to the unique nature of the training set.Broader legal, social, and ethical issues were also noted in the study, such as the ethical application of Artificial Intelligence (AI), adherence to the UK GDPR, and fair access to digital tools for small holder farmers.This research successfully demonstrated the feasibility and effectiveness of the developed lightweight deep learning model through the performance results which showed its capacity in classifying watermelon diseases within limited resources like mobile devices when compared to heavyweight models like VGG16. This becomes significant for AI solution for disease management in a low resource environment for sustainable agriculture. Future research could focus on enhancing model interpretability, increasing robustness and generalisability of the model through expansion and integration of realistic field data during model training and diversification of model test on different low resource devices.
Keywords: Artificial Intelligence, anthracnose, crop disease, downy mildew, mosaic virus, sustainable agriculture, watermelon
