Doing more with less: A comparative assessment between morphometric indices and machine learning models for automated gully pattern extraction (A case study: Dashtiari region, Sistan and Baluchestan Province)

Abstract

Deep gullies in the Dashtiari Region prompted us to couple different morphometric indices obtained from a UAV-derived DEM to automatically extract gully signatures. The extraction of gully signatures is commonly undertaken via pattern recognition techniques, whose recent advancements seem to require more data and rather cumbersome modeling processes, making it even more difficult for those who are not well-versed in such contexts. Among these methods, object-based image analysis (OBIA), machine learning, and deep learning techniques are the most common. Conversely, here we took advantage of simple morphometric indices and their combinations for gully extraction, including valley depth (VD), topographic position index (TPI), positive openness (PO), red relief image map (RRIM), elevation, slope degree, and the coupled PO-DEM. Furthermore, we compared the automatically derived gully patterns to the manually extracted ones (treated as the ground truth), and their spatial autocorrelation was investigated. Additionally, the application of the classification tree (CT) as a powerful machine learning model was comparatively assessed for morphometric indices. The performance of the adopted pattern extraction techniques was estimated using four different metrics: precision index, true skill statistics (TSS), Cohen's kappa, and Matthews correlation coefficient (MCC). The results revealed that the single use of PO, TPI, and RRIM indices failed to reliably capture the gullies’ pattern, leading to partial success. Notably, combinations of indices showed that the coupled PO-DEM could successfully classify the gully presence locations from the absences and outperform the CT model in terms of both goodness-of-fit and generalization capacity (prediction power), considering all four-performance metrics. Hence, comparing the amount of time spent for manual delineation of gullies, the application of simple morphometric indices, and machine learning models is beyond comparison.