Abstract
Soil-borne pathogens such as Fusarium oxysporum are responsible for substantial yield losses in high-value crops, yet most existing digital plant disease detection systems remain reactive, relying on post-infection visual symptoms that emerge too late for effective intervention. While Internet of Things (IoT) technologies and Machine Learning (ML) have been widely adopted for image-based disease diagnosis, their application to pre-planting, soil-level risk assessment remains limited. This study proposes an integrated IoT–ML framework for pre-planting Fusarium wilt risk assessment, leveraging a hybrid feature set that fuses real-time soil sensor data with static farmer-provided soil characteristics. An ESP32-based sensing platform was deployed to collect soil pH, temperature, and moisture, which were combined with manually entered soil texture (sand and clay fractions) and nutrient indicators (NPK). Fifteen supervised Machine Learning models were evaluated to assess predictive reliability under noisy, field-relevant conditions. Among all models, the Bagging Classifier achieved the highest performance, with an accuracy of 92.1% and a Cohen’s Kappa of 0.88, indicating strong agreement beyond chance. In contrast, gradient boosting methods such as XGBoost achieved only 71% accuracy, highlighting their sensitivity to sensor noise and domain variability. Results demonstrate that variance-reduction ensembles, combined with hybrid data fusion, are particularly well-suited for soil-based disease risk modelling. By enabling actionable risk assessment before sowing, the proposed system supports proactive soil management strategies, allowing farmers to implement mitigation measures prior to crop establishment. This work advances the shift from reactive disease detection toward soil-centric, decision-oriented precision agriculture.