Mapping permanent gullies in an agricultural area using satellite images: Efficacy of machine learning algorithms

Abstract

Gullies are responsible for detaching massive volumes of productive soil, dissecting natural landscape and causing damages to infrastructure. Despite existing research, the gravity of the gully erosion problem underscores the urgent need for accurate mapping of gullies, a first but essential step toward sustainable management of soil resources. This study aims to obtain the spatial distribution of gullies through comparing various classifiers: k-dimensional tree K-Nearest Neighbor (k-d tree KNN), Minimum Distance (MD), Maximum Likelihood (ML), and Random Forest (RF). Results indicated that all the classifiers, with the exception of ML, achieved an overall accuracy (OA) of at least 0.85. RF had the highest OA (0.94), although it was outperformed in gully identification by MD (0% commission), but the omission error was 20% (MD). Accordingly, RF was considered as the best algorithm, having 13% error in both adding (commission) and omitting pixels as gullies. Thus, RF ensured a reliable outcome to map the spatial distribution of gullies. RF-derived gully density map reflected the agricultural areas most exposed to gully erosion. Our approach of using satellite imagery has certain limitations, and can be used only in arid or semiarid regions where gullies are not covered by dense vegetation as the vegetation biases the extracted gullies. The approach also provides a solution to the lack of laser scanned data, especially in the context of the study area, providing better accuracy and wider application possibilities.