Comparative Analysis of Machine Learning Algorithms for Soil Erosion Modelling Based on Remotely Sensed Data

Abstract

Recent years have seen an increase in the use of remote-sensing based methods to assess soil erosion, mainly due to the availability of freely accessible satellite data, with successful results on a consistent basis. There would be valuable benefits from applying these techniques to the Arctic areas, where ground local studies are typically difficult to perform due to hardly accessible roads and lands. At the same time, however, the application of remote-sensing methods comes with its own set of challenges when it comes to the peculiar features of the Arctic: short growing periods, winter storms, wind, and frequent cloud and snow cover. In this study we perform a comparative analysis of three commonly used classification algorithms: Support Vector Machine (SVM), Random Forest (RF) and Multilayer Perceptron (MLP), in combination with ground truth samples from regions all over Iceland, provided by Iceland’s Soil Conservation Service department. The process can be automated to predict soil erosion risk for larger, less accessible areas from Sentinel-2 images. The analysis performed on validation data sets supports the effectiveness of both approaches for modeling soil erosion, albeit differences are highlighted.

The Effect of Bioclimatic Covariates on Ensemble Machine Learning Prediction of Total Soil Carbon in the Pannonian Biogeoregion

“…Moreover, the optimal C value (C = 0.25) emphasized a simpler decision boundary with controlled margins, so the model exhibited improved generalization capabilities. This well-balanced regularization approach helped avoid overfitting, contributing to enhanced model robustness [46].…”

A Review of Machine Learning Techniques in Agroclimatic Studies

“…Agricultural ML tasks typically categorize into supervised or unsupervised learning [19,32,37]. Supervised learning, employing methods like Gradient Boosting (GB) and neural networks (NNs), suits predictive modeling tasks, such as estimating crop yields.…”

“…Several ML and DL studies tend to focus on the outcomes rather than the journey of reaching those outcomes [19][20][21]37]. For instance, studies often highlight the accuracy and effectiveness of their models without providing a comprehensive breakdown of the model architecture, parameter settings, or training processes [18,20,28].…”

Predicting Soil Erosion Using RUSLE and GeoSOS-FLUS Models: A Case Study in Kunming, China

“…The study is based on the R language platform, with tuned parameters: the number of decision trees in the model (ntree) = 750 and the number of variables tried at each node (mtry) = 5, with other settings as default. The accuracy of the model is validated by the coefficient of determination (R 2 ) and mean absolute error (MAE); a higher R 2 and a lower MAE indicate a higher explanatory accuracy of the model [47][48][49] Woodland and grassland were the main land use types in Kunming, widely distributed and accounting for more than 70% of the study area (Figure 2). Cropland was the next major type, covering more than 20% of the study area, primarily distributed in low-altitude regions such as Songming and Yiliang.…”