Spectroscopic Sensor Data Fusion to Improve the Prediction of Soil Nutrient Contents

This study aims to advance the understanding and application of spectroscopic sensor data fusion for improving soil nutrient content predictions. In addition to presenting an extensive review of studies on the spectroscopic sensor data fusion, a research investigation was conducted to assess the effectiveness of five fusion algorithms in predicting three primary nutrients (Nitrogen, Phosphorous, and Potassium) and two secondary nutrients (Calcium and Magnesium) in soil using Visible and Near-Infrared, Mid-Infrared, and X-ray Fluorescence data. Among the five fusion algorithms, one was a low-level fusion involving data concatenation. Two were mid-level fusions, incorporating feature extraction by applying (i) Principal Component reduction and (ii) Partial Least Squares reduction. The other two were high-level fusions, namely (i) Simple Averaging and (ii) Granger Ramanathan Averaging. The results indicate that Low-Level Fusion may not be suitable for inherently incompatible data. Mid-level fusion improved the R² by 0.1-18%, RMSE by 0-8%, and RPIQ by 0-11.5%, while high-level fusion enhanced the R² by 0-12.5%, RMSE by 0-12%, and RPIQ by 2.3-13.4%, depending on the nutrients and fusion algorithms. Despite these improvements, predictions were only satisfactory for primary nutrients, and none of the algorithms could notably enhance predictions for Phosphorous. The study also finds that fusion algorithms do not significantly improve bias. The study provided evidence of improvement in prediction accuracy with data fusion which can aid in delineating management zones for precision agriculture. It also encourages further research on novel approaches of sensor fusion and algorithms that can effectively handle non-linearity introduced due to the fusion of data. Advisor: Yufeng Ge