A Hierarchical Deep Learning Framework for Robust Cassava Crop Verification and Disease Diagnosis

Abstract Cassava represents a crucial staple crop in sub-Saharan Africa, whose productivity is severely threatened by viral and bacterial diseases. While deep learning approaches show promise for automated disease detection, they often operate under unrealistic assumptions and suffer from dataset limitations including class imbalance and label noise. This paper presents a novel hierarchical deep learning framework that addresses these challenges through a three-stage pipeline: crop verification, health assessment, and disease classification. Our approach achieves computational efficiency through conditional execution while maintaining diagnostic accuracy. On the Cassava Leaf Disease Classification dataset comprising 21,397 images, our framework achieves 91.8\% overall accuracy with statistically significant improvements over flat classification baselines (mean difference: 1.7\%, bootstrap 95\% CI: 1.2-2.2\%, p < 0.001, paired t-test on 5-fold cross-validation). The framework demonstrates robust performance across all disease classes, with particular strength in handling class imbalance through a multi-focal loss formulation. Comprehensive validation includes per-class metrics, 5-fold cross-validation, ablation studies, explainability analysis, and out-of-distribution evaluation. Our work provides a validated blueprint for hierarchical approaches in agricultural AI, with complete reproducibility through public code and dataset splits.