Prediction of Total Soil Nitrogen Variations Using Three Machine Learning Approaches and Remote Sensing Data

Every country relies on soil as a vital natural resource that significantly contributes to environmental conservation and food production. Preparation of soil nutrient distribution map serves as a valuable tool for managers to make decisions. Due to the time-consuming and expensive nature of laboratory analysis for these variables on a large scale, efforts have been made to explore soil nitrogen through remote sensing. The current research deals with the application of remote sensing methods along with regression and random forest models to predict total soil nitrogen in Gilan province. This study aimed to answer two main questions: (1) Can SAR data be used to quantify total soil nitrogen (TSN) (2) How do SVM, BRT and RF algorithms perform in predicting soil nitrogen content?

This study focused on evaluating the data capabilities of Landsat-9 and Sentinel-1 satellites individually and in combination, using advanced algorithms such as Support Vector Machine (SVM), Boosted Regression Tree (BRT), and Random Forest (RF). The purpose of this evaluation was strategic, aiming to showcase the diverse conditions of the study area based on land cover/land use, climatic, and topographical parameters. Various variables, including climate parameters, topographic components, and remote sensing subscale indices, were investigated in conjunction with SAR data and optical images. Nonlinear machine learning algorithms, specifically SVM, RF, and BRT, were employed to predict total soil nitrogen status by modeling complex relationships between soil properties and environmental variables. R software, utilizing the CARET package for parameter input, was employed to implement the algorithm.

The results indicated the following: RF and BRT algorithms outperformed SVM and were effective in monitoring total soil nitrogen values. Multi-temporal SAR images showed higher accuracy in monitoring total soil nitrogen content compared to optical remote sensing data, facilitating more realistic predictions in paddy soils. The integration of environmental variables led to an increase in the accuracy of algorithms, where remote sensing variables played a crucial role, contributing to 61% and 51% effects in RF and BRT algorithms, respectively. The comparison of SVM and RF algorithms revealed that RF ranked second after the BRT algorithm, and the accuracy of total soil nitrogen estimation was not achieved with the SVM algorithm. However, both BRT and RF algorithms were able to monitor changes in total soil nitrogen. BRT performed better, accurately recording 58% of changes, as evidenced by a higher R2 value (0.58) and lower RMSE (0.25 mg/kg) and MAE (0.19 mg/kg) values.

In conclusion, the following key points were extracted from this research: 1) RF and BRT algorithms outperformed SVM in effectively monitoring total soil nitrogen levels; 2) multi-temporal SAR images demonstrated higher accuracy in tracking total soil nitrogen compared to optical remote sensing, enabling precise predictions in paddy soils; 3) the incorporation of environmental variables enhanced algorithmic accuracy; and 4) remote sensing variables contributed 61% and 51% to RF and BRT algorithms, respectively.