Crop diseases threaten to trigger cascading failures across global food systems, devastating ecosystems, economies and livelihoods. In such historical agricultural collapses, smallholder farmers in low-resource settings have been disproportionately harmed. As agricultural technology becomes democratised, we see new opportunities to develop tools that promote stability and incremental improvement. While numerous Artificial intelligence (AI) and Machine learning (ML) models have been developed for plant disease detection, most rely on generic architectures trained on limited or synthetic datasets. As such, they struggle to detect early or non-visible symptoms, and don’t run efficiently on low-power hardware. This thesis addresses early disease detection in cocoa crops by integrating advanced computer vision (CV) techniques with biological knowledge, offering effective, efficient, and transparent AI models tailored to resource-constrained environments. Key contributions include Phyt-Net, a custom lightweight convolutional neural network (CNN) architecture designed for accurate diagnosis with limited hardware and training data. We investigate spectroscopy and infrared (IR) imaging as sources of complementary signal beyond the visible spectrum, finding infrared imagery can be competitive in image-based classification. We leverage semi-supervised learning and a novel dynamic focal loss (DFLoss) to direct model attention toward difficult-to-detect symptoms, and enhance interpretability through gradient-weighted class activation mapping (Grad-CAM) visualisations, enabling validation of model focus against biological symptoms. We also present a new high-quality benchmark dataset of 7,220 images of diseased and healthy cocoa trees, offering a greater and more realistic challenge than existing benchmarks like PlantVillage. Extensive experiments demonstrate that tailored architecture design and advanced training procedures improve detection accuracy, generalisation and resource efficiency, while non-visible sensing analyses narrow where complementary signal is most plausibly useful. By bridging computer science, plant pathology, and exploratory non-visible sensing, this work yields new methods for AI-driven disease surveillance tools that can help smallholder farmers protect crops and promote sustainable agriculture, highlighting the societal and ecological relevance of such interdisciplinary work.
