Optimizing machine learning algorithms for spatial prediction of gully erosion susceptibility with four training...

Area of Science:

• Environmental Science
• Geosciences
• Remote Sensing

Background:

• Gully erosion significantly impacts agricultural sustainability worldwide.
• Understanding gully erosion mechanisms requires accurate mapping.
• Land modification and damage to agricultural fields are key concerns.

Purpose of the Study:

• To develop a new modeling approach for gully erosion susceptibility mapping (GESM).
• To assess the performance of machine learning models with varying data sample sizes.
• To identify key factors influencing gully erosion in the Golestan Dam basin, Iran.

Main Methods:

• Compiled a spatial database of 14 gully erosion (GE) factors at 1042 locations.
• Utilized four machine learning models: maximum entropy (MaxEnt), general linear model (GLM), support vector machine (SVM), and artificial neural network (ANN).
• Employed four training dataset scenarios (100%, 75%, 50%, 25%) and validated using the receiver operating characteristic (ROC) curve.

Main Results:

• Random Forest analysis identified distance from stream, elevation, distance from road, and vertical distance of the channel network (VDCN) as critical factors.
• Sample size significantly influenced the performance of most machine learning algorithms.
• Artificial neural network (ANN) demonstrated the best performance (AUROC 85.7-90.4%) with minimal sensitivity to sample size reduction.

Conclusions:

• ANN is a robust model for GESM, performing well even with limited data.
• The study provides valuable insights for selecting appropriate machine learning methods when complete gully inventories are unavailable.
• Findings support improved land management strategies to mitigate gully erosion impacts on agriculture.