A Machine Learning Approach to Estimating Land Use Change and Scenario Influence in Soil Infiltration at The Sub-Watershed Level

Abstract

This research uses random forests to develop infiltration-friendly land use scenarios, addressing the global 32% change in land use over the past six decades. The study used Sentinel-2A for 2017, 2019, 2021, and 2022 as a land use baseline, predicting business as usual using cellular automata and comparing it with regional spatial planning and land capability scenarios. One hundred points of infiltration data were distributed using a random forest. Results showed that deforestation and its change into orchards, rice fields, and settlements over five years affected the infiltration. Business as usual reduces the high infiltration class to approximately 1,545 ha, while regional spatial planning and land capability cover 1,390 ha and 1,316 ha, respectively. The most infiltration-friendly land use scenario is applicable at the sub-watershed level, with an accuracy of about 97%. This research limitations include not comparing extreme dry seasons and using 2022 infiltration values for all other years.

Keywords

• machine learning; remote sensing; geostatistic; hydrology; disaster.