Soil type classification using Landsat 8: A comparison between the USDA and a local system in Israel

Digital Soil Mapping (DSM) is an essential tool for understanding the complex relationship between soil and the environment. In this study, we digitized the soil map of Israel created by Ravikovitch in 1969 (that was based on a local classification system) and used Landsat 8 spectral data to predict soil classes across Israel using machine learning. We also made a similar analysis using a global USDA soil orders layer. This work is pioneering, and this is the first attempt to transfer the enormous and important work done by Ravikovitch to the digital level by combining this map with satellite observations of Landsat 8. Our study showed that the spectral-based predictions using Landsat 8 data in combination with the USDA soil orders data and machine learning techniques resulted in very accurate predictions of USDA soil orders in Israel (accuracy = 0.84) and in Cyprus (accuracy = 0.88). We also tested the transferability of the Israeli USDA soil orders model to Cyprus, a nearby country with a similar soil taxonomy, however, poor accuracies were obtained at this stage (accuracy = 0.13). The predictions on the digital map of Ravikovitch were intermediate (accuracy = 0.54) because so many classes were required to predict (24 classes). Our study highlights the importance of digitizing and updating existing soil maps, and demonstrates the potential of combining machine learning with satellite spectral data for accurate soil classification.