Prediction of total soil nitrogen variations using three machine learning approaches and remote sensing data

Abstract

Abstract The fluctuation of total soil nitrogen (TSN) levels, whether it be an excessive increase or decrease, can result in microbial contamination, decreased vegetation coverage, and reduced agricultural product yield. However, analyzing nitrogen levels in a laboratory setting can be a costly and time-consuming process when done on a large scale. As a solution, remote sensing technology can be utilized to address this issue. In this research, the data capabilities of Landsat-9 and Sentinel-1 satellites and their integration along, with the use of support vector machine (SVM), boosted regression tree (BRT), and random forest (RF) algorithms, were evaluated in the zoning of TSN values in the soil of paddy fields in northern Iran. Several variables were used that had the potential to predict TSN values. TSN estimation accuracy was not achieved with the SVM algorithm. However, the BRT and RF algorithms were able to monitor TSN changes, with the BRT performing better by accurately capturing 58% of changes due to its higher R2 value (0.58) and lower RMSE (0.25) and MAE (0.19) values. LULC maps and BC-3 band data variables play a key role in producing the TSN map. Hence, utilizing SAR data in conjunction with machine learning algorithms is a viable recommendation for monitoring soil nitrogen levels, particularly in regions with high rainfall where the sky is frequently overcast with clouds and fog. SAR data, with the biggest share (31%), was the most important variable in the BRT algorithm.