Machine Learning and Fog Computing-Enabled Sensor Drift Management in Precision Agriculture

Despite tremendous improvements in sensing mechanisms at the hardware level, sensor drift is still an issue for reliable Internet of Things (IoT) applications. Existing literature is limited in exploring the impacts of sensing contexts on sensor drift and handling sensor drift using soft computing approaches. The study intends to propose a fog-enabled IoT architecture to explore the impacts of different sensing contexts on sensor drift by sampling in remote areas without any connectivity issues. The proposed solution also incorporates machine learning capabilities for sensor drift-level detection. The proposed solution is implemented for drift management of soil fertility sensors for sensing soil nitrogen (N) levels at different levels of soil electric conductivity (EC), soil pH, soil moisture, and soil temperature. The sensor drift is observed by comparing the nitrogen (N) sensed values against the standard method for observing soil nitrogen (N) levels. The evaluation of various machine learning models for sensor drift detection reveals that light gradient boosting machine regression (LGBMR) outperforms other models, demonstrating superior predictive capabilities with a high mean coefficient of determination ( $R<^>{{2}}$ ) of 0.87 and lower error metrics across fivefold cross validation. The application of the proposed solution in the real world demonstrates that nitrogen (N)-level observations by the proposed solution are more accurate with a mean difference of 1.43 mg/kg against the standard method. The proposed solution has several implications for IoT applications in precision and smart agriculture.