A Critical Comparative Survey of ML/DL Approaches for Late Blight Detection and Prediction for Guinean Agriculture

Abstract

Late blight (Phytophthora infestans) remains one of the most devastating diseases for staple crops. This survey benchmarks CNN (AlexNet, VGG-16, ResNet-50, DenseNet) and RNN architectures (LSTM, GRU) and hybrid multimodal frameworks. Our analysis reveals that ResNet-50 offers the best accuracy-to-parameter trade-off for image detection (90-99 %), while LSTM outperforms GRU for long-term temporal dependencies in disease risk prediction. We identify four structural limitations: over-reliance on the PlantVillage benchmark, near-total absence of West African training data (the rare African studies (Tanzania, Cameroon) confirm a $\mathbf{3 2 \%} \mathbf{F 1}$-score drop in real-world conditions), neglect of deployment constraints for resource-limited environments, and insufficient integration of image and climate modalities. Based on this analysis, we propose a hybrid CNN-LSTM reference architecture and formulate concrete recommendations for its adaptation to the Guinean agricultural context.