A Systematic Review of Artificial Intelligence and Deep Learning Approaches for Cocoa Pod Disease Detection

Cocoa pod diseases destroy an estimated 700,000 metric tons of cocoa annually, threatening 40–50 million smallholder farmers across tropical regions. Accurate field diagnosis remains constrained by symptom similarity, labour-intensive scouting, and limited specialist access in remote communities. This paper presents a PRISMA-guided systematic review of 25 peer-reviewed studies on automated cocoa pod disease detection and classification (2016–2026), addressing architectural approaches, implementation strategies, performance outcomes, and research gaps. Studies were identified through a multi-database search (Google Scholar, IEEE Xplore, ScienceDirect, SpringerLink, PubMed) and appraised across seven quality dimensions including dataset adequacy, methodological rigour, evaluation completeness, and deployment feasibility. Convolutional neural networks dominate the literature (≈44% of algorithmic instances), with transfer learning from EfficientNet, MobileNet, ResNet, and VGG consistently outperforming training from scratch. Object detection frameworks (YOLO variants, SSD) account for 27% of instances, reflecting a shift toward spatial localisation. The Learnable Gated Fusion Convolutional Block Attention Module (LGF-CBAM), integrated with ResNetV2-101, achieves 98.95% classification accuracy, the strongest result validated across five independent datasets. Despite accuracy gains, six structural deficiencies undermine confidence in reported findings: Near-universal absence of statistical significance testing; heterogeneous evaluation metrics; no ablation studies; exclusive reliance on held-out test sets; inadequate class imbalance handling; and underreporting of computational complexity. Approximately 32% of studies report no overfitting mitigation, fewer than 20% employ cross-validation, 69% include no mobile or edge deployment component, and 32% use fewer than 500 training images. No current model achieves balanced strength across accuracy, dataset adequacy, overfitting control, mobile suitability, and computational efficiency. The review recommends: Attention-augmented and hybrid CNN-ViT architectures with integrated explainability; coordinated development of large-scale, georeferenced benchmark datasets; standardized evaluation protocols mandating cross-validation and full metric reporting; and deployment-oriented design prioritizing on-device inference, offline functionality, and formal usability testing with farmer populations.