A semi-automatic approach to derive land cover classification in soil loss models

Soil erosion constitute a major threat to human lives and assets worldwide, as well as a major environmental disturbance. The Revised Universal Soil Loss Equation (RUSLE) integrated with Geographical Information System (GIS) has been the most widely used model in predicting and mapping soil erosion loss. Remote sensing has particular utility for soil loss model applications, providing observations on several key aspects of Land use and Land cover (LULC) linked to the cover-management factor C of the RUSLE, over wide areas and in consistent and repeatable measurements. A free and open source GIS application coupled with remote sensing data was developed under QGIS software allowing to improve the C factor functionality: (i) automatically download satellite images; (ii) clip with the study case and; (ii) perform a supervised or unsupervised classification, in order to obtain the land cover classification and produce the final C map. One of the most efficient supervised classification algorithms is the Support Vector Machine (SVM). Random Forest (RF) is also an easy-to-use machine learning algorithm for supervised classification. The automation of this functionality was based in the R and SAGA software, both integrated in QGIS. To perform the supervised classification, SVM and RF methods were incorporated. The overall accuracy and Kappa values are also automatically obtained by the R script and GRASS algorithms, which allows to evaluate the result obtained. To perform the unsupervised classification K-means algorithm from SAGA was used. This updating in RUSLE application improve the results obtained for C factor and help us to obtain a most accurate estimation of RUSLE erosion risk map. The application was tested using Sentinel 2A images in two different periods, after and before the forest fire event in Coimbra region, Portugal. In the end, the three resulted maps from SVM, RF and K-means classification were compared.