<![CDATA[This study implements a lightweight TinyML model to classify soil moisture conditions and support irrigation decisions in rice cultivation, chosen over conventional cloud-based ML because it enables low-power, low-latency, fully offline inference on microcontrollers—critical for rural areas with limited connectivity. Trained on 3,021 localized microclimate records from Denai Lama Village (temperature, humidity, rainfall, cloud cover) using logistic regression for its simplicity and interpretability under resource constraints, the model was deployed on an ESP32 for real-time predictions into three classes (underwatered, optimal, overwatered). Experimental results show accuracy = 0.982 and weighted F1 = 0.982 on the validation set (ROC–AUC = 0.997), and on the held-out test set (N = 194) the model achieved 93.4% accuracy, 0.927 weighted F1 (precision 0.914; recall 0.942), and ROC–AUC = 0.988. These findings indicate that TinyML provides a practical, low-cost, and scalable edge-AI pathway for reliable, energy-efficient decision support in precision irrigation without network dependence, offering a deployable template for smallholder farming contexts.]]>
Copyrights © 2025
