This is the authors' abstract. We don't add key points for this paper.
Background Tobacco is Zimbabwe’s most economically significant agricultural export commodity, generating annual revenues exceeding USD 1 billion and sustaining livelihoods for hundreds of thousands of smallholder farmers. Farm management decisions remain predominantly reactive, relying on visual inspection and experiential judgement rather than data-driven predictive intelligence. Objective This paper presents SADSS-Tobacco (Smart Artificial Intelligence Driven Decision Support System for Tobacco), an integrated multi-modal framework combining deep learning, Internet of Things (IoT) environmental sensing, and explainable artificial intelligence (XAI) for leaf-stress detection, irrigation forecasting, yield prediction, and curing optimisation in Zimbabwean tobacco production. The leaf-stress component is a three-class severity detector (healthy, early-stage, advanced-stage); it is intended as a severity alert to prioritise inspection, not as a differential diagnosis of the underlying cause (nutrient deficiency, drought, or pathogen). Methods A ResNet50 convolutional neural network was fine-tuned on 1,152 field-collected, expert-labelled tobacco leaf images (Fleiss’ κ = 0.87; Eq. 6) for three-class stress classification. A Random Forest regression model was trained on approximately 50,000 IoT sensor readings for 24-hour irrigation forecasting. An XGBoost regression model was developed from 48 field-year TIMB historical records (2018–2025) for seasonal yield prediction. Explainability was implemented via Grad-CAM (image models) and SHAP (Eq. 4, tabular models). Five-fold cross-validation and a structured ablation study were applied throughout. Results ResNet50 achieved 92.3% accuracy (95% CI: 89.1–94.8%; F1-macro = 0.91; Brier = 0.091). The irrigation model attained RMSE = 0.12 m³/m³ (R² = 0.81). On a small independent 2025 test set (n = 6 field-year observations), yield forecasting attained RMSE = 142 kg/ha (Eq. 3) and R² = 0.78 (MAPE = 7.2%) and revenue-grade prediction reached 84% accuracy (Wilson-score 95% CI: 43–97%); owing to this limited cohort, both are reported as preliminary pilot results rather than validated performance. Multi-modal fusion improved F1-macro by 12.3% over the image-only baseline (Cohen’s d = 1.04; p < 0.001, Bonferroni-corrected). Conclusions This pilot indicates that integrating multi-source data through explainable deep learning can support a shift from reactive to proactive, evidence-based tobacco management, offering a replicable architecture for export-oriented cash crops in data-scarce Sub-Saharan African environments; full validation awaits a planned 2026–2027 randomised controlled trial.
Answers are grounded in this paper and cite their sources.
No takes yet. Share an insight, caveat, or question.
Jongwe et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: