ExMAD (Expert-based Multitemporal AI Detector): An open-source methodological framework for remote and field landslide inventory

Landslides threaten lives and infrastructure, making accurate inventories crucial for risk management. This study combines expert methods with machine learning to automate and validate landslide detection and timing using Sentinel-2 satellite imagery. We developed ExMAD (Expert-based Multi-temporal AI Detector), an open-source methodological framework ( https://github.com/NewGeoProjects/ExMAD ) to integrate artificial intelligence with human expertise to detect occurrence timing of a targeted landslide. A U-Net neural network was chosen to effectively test ExMAD in landslide detection over Sentinel-2 worldwide multitemporal satellite imagery sequences, and the model was tested through five evaluations. ExMAD was able to effectively extract timing of target landslides on Sentinel-2 images and was able to correctly detect the presence/absence of landslide, proving the suitability of AI systems in landslide temporal mapping task. This research proves the potential of hybrid AI-human approaches for landslide risk assessment, integrate human expertise with machine learning offers promising advancements for remote and field mapping of landslide. Furthermore, the ExMAD methodology adheres to the European Union's Artificial Intelligence Act, stressing human oversight in high-risk AI applications to enhance trust, control, and efficiency in landslide inventory creation and risk management. • ExMAD integrates human expertise with AI for landslide inventory creation process. • The framework uses SENTINEL-2 satellite data for worldwide long-term applicability. • A trained U-Net model accurately segments shallow landslides. • An open-source U-Net model example is provided under the MIT license. • Hybrid approach introduces human oversight in risk management using AI systems.