Mapping Soil Parameters with Environmental Covariates and Land Cover Projection in Tropical Rainforest, Hangadi Watershed, Ethiopia

Abstract

Machine learning and geostatistics are efficient techniques for investigating the geographic distribution of soil properties. This study’s objectives were to assess soil fertility status, map the spatial variability of selected soil parameters and compare random forest with ordinary kriging. Soil samples were collected to analyze parameters: pH, cation exchange capacity (CEC) and organic carbon (OC) using systematic sampling. Some environmental covariates were used in the machine learning process: a digital elevation model (DEM) collected from USGS distributing shuttle radar topography mission data and a LULC map generated from a 30-year time series (1988–2018) of Landsat 8. Georeferenced samples were sent to Batu Soil Research Laboratory. pH, CEC and OC were mapped and status was determined using random forest and ordinary kriging. Random forest was more accurate with low mean square error (MSE), root mean square error (RMSE), mean absolute error (MAE) and coefficient of determination (high R2). In random forest, pH varied between 5.03 and 5.76 and ordinary kriging pH ranged from 4.96 to 5.76. pH was greater in cultivated land. CEC and OC were higher in the forest. The higher pH in cultivated land was due to grass coverage and minimal tillage. The addition of organic matter and CEC to a forest may result in higher OC. Environmental covariates (topographic, bands, NDVI and LULC) were used to predict the gradients of pH, OC and CEC. For pH, OC and CEC, DEM was the most important predictor. CEC was high in low landscape, but low in high landscape positions. Low OC requires composting, fallow and organic fertilizers. Future research should include the remaining predictors: physiochemical and lithological data to improve the performance of random forest.