Testing NDVI and U-Net for automated mapping of multiple-occurrence regional landslide events using satellite and aerial multispectral data (Casola Valsenio, Emilia-Romagna, Northern Apennines, Italy)

In May 2023, the Emilia-Romagna Region (northern Italy) experienced a severe multiple occurrence regional landslide event, resulting in more than 80,000 debris slides and flows as well as some hundreds of rock-block slides. One of the more severely affected areas was the Municipality of Casola Valsenio, with more than 5200 landslides in 84 km2. In this study, we assess the performances of two established methods for automated landslide mapping (NDVI change and U-Net) applied to satellite and aerial multispectral data of resolutions ranging from 10 to 0.2 m (the latter resampled to 1 and 2 m). In doing so, we used a basic configuration (i.e., no additional data layer in the analysis) to simulate their application in the aftermath of a MORLE scenario, when time and training data are limited. The results show that, even with limited training, U-Net outperforms NDVI in terms of accuracy only in case of 2 m resolution aerial data (achieving an F1-score higher than 0.6), while with 10 m resolution satellite data, the NDVI performs better (F1 up to 0.5) than U-Net (F1 lower than 0.4). The enhanced performance of U-Net with higher resolution data is ascribed to its proficiency in detecting landslides not only by changes in vegetation but also by analyzing the landslide's physical shape. However, the research ultimately evidence that in the case study area, automated mapping using NDVI change and U-Net in basic configuration still is affected by too many false positives and misses, so it cannot eliminate the need for manual verification, especially for precise identification of affected buildings and roads.