Spatial Modeling of Soil 137Cs Inventory and Redistribution Using Machine Learning with Remote Sensing and Auxiliary Variables at the Watershed Scale

The radionuclide 137 Cs has been widely used as a reliable tool for estimating soil erosion at hillslope and watershed scales. However, measuring 137 Cs inventory at the watershed scale is both costly and time-consuming. To address this challenge, spatial models have been developed to improve the accuracy of soil loss and deposition estimations. This study focused on predicting 137 Cs inventory and its redistribution across a watershed using machine learning algorithms and environmental variables. A total of 100 soil samples were collected from the Aji-Chai watershed in northwestern Iran, covering depths of 0–15 and 15-30 cm, and analyzed for 137 Cs inventory. Four modeling scenarios were evaluated to identify the most influential environmental variables: remote sensing (RS) data (S1), topographic attributes (S2), a combination of RS and topographic attributes (S3), and a comprehensive set of variables, including geology and land use maps (S4). Among these, the Random Forest (RF) model performed best in scenario S4, achieving an R 2 of 0.83, a CCC of 0.84, and an nRMSE of 0.20; and an R 2 of 0.78, a CCC of 0.80, and an nRMSE of –0.31 for estimating 137 Cs inventory and 137 Cs loss/deposition , respectively. Variable importance analysis revealed that topographic attributs such as diffuse insolation (Diffuse INS), channel network base level (CNBL), valley depth channel (VDC), and the LS-Factor, played a critical in explaining the spatial distribution of 137 Cs and 137 Cs loss/deposition . The spatial modeling approach provides valuable insights into erosion and sedimentation patterns within the watershed, facilitating the development of effective soil conservation and land management strategies to promote sustainable land use practices.