Landslide Detection Using Swin Transformer Model and PlanetScope Images: A Performance Comparison by the Input Data Combination

Landslide areas have the limitation of less distinct spectral characteristics compared to other land cover types.To address this, combining spectral information provided by satellite imagery with other types of information can improve the performance of deep learning-based landslide detection.This study analyzes the performance differences of a deep learning model when various combinations of spectral information, vegetation information, topographic information, and texture information are used.The model used in this study is the Swin Transformer, which is known for its high performance in landslide detection, along with PlanetScope satellite data that provides daily global observations.The experimental results showed that using only spectral information achieved a mean Intersection over Union (mIoU) accuracy of 72.08%, while combining spectral, vegetation, terrain, and texture information in various ways resulted in a maximum mIoU accuracy of 74.31%.Although this demonstrates an improvement in model performance, the difference was not substantial.When the model was applied to test datasets, performance differences were observed depending on regional characteristics.In regions with extensive forest coverage, datasets utilizing diverse types of information exhibited improved predictive capabilities compared to those using only spectral information.However, in regions where forests, croplands, bare land, and developed areas were intermixed, the model's relative detection ability tended to decrease.Even in these cases, the combination of spectral, vegetation, topographic, and texture information resulted in relatively superior detection capabilities.This study highlights the importance of carefully selecting input information in landslide detection using deep learning models and satellite imagery, demonstrating how such choices significantly influence model performance.