Influencia de la heterogeneidad espacial en la evaluación de la susceptibilidad por movimientos en masa

This study explores the incorporation of spatial heterogeneity in modeling landslide susceptibility in the complex terrain of the Colombian Andes. The data used included a dataset of 13,777 landslides recorded between 1970 and 2023, derived from high-resolution optical imagery, as well as precipitation data from the CHIRPS dataset. Additional spatial information, such as land cover, was derived using Sentinel-2 data processed in Google Earth Engine. These diverse geospatial data sources were integrated to provide a comprehensive understanding of the factors influencing landslide susceptibility. By applying Geographically Weighted Regression (GWR) and its extension, Multiband Geographically Weighted Regression (MGWR), we demonstrate significant improvements in the capacity of models to capture local variability and enhance predictive accuracy. GWR and MGWR allow for the identification of spatially differentiated patterns and specific scales of influence for each predictor, leading to a more precise representation of landslide susceptibility and highlighting the necessity of adapting models to the unique geographic context of each region. However, the study also underscores the need for careful bandwidth selection to avoid overfitting and ensure model stability. Despite the advantages of GWR and MGWR, limitations such as high computational complexity and local collinearity were identified, which can impact the interpretation and reliability of the result.