Mapping of landslides by using partially cloudy optical Earth observation imagery: a case study of 2023 Türkiye Earthquakes

Landslide mapping is essential for hazard assessment and disaster response, and methods based on Earth observation (EO) enable the mapping of large areas impacted by major disasters.  These methods, however, often rely on cloud-free optical images, which are rarely available in high-rainfall areas prone to landslides, delaying timely detection. Furthermore, mosaicking multiple consecutive images to eliminate cloud cover discards valuable temporal information, such as the actual timing of landslide occurrences and the progression of their extents over time. 

To address these challenges, we introduce a novel method that processes successive partially cloudy images to detect visible landslide extents and automatically aggregates this information for rapid first detection and accurate spatiotemporal mapping of landslides. The model-agnostic method supports various EO-based landslide detection models from the literature. It uses binary model outputs (landslide / no landslide), associated uncertainty levels (if available), and cloud mask data together with cloud uncertainty to classify individual image cells into four states: landslide, background, unknown (e.g., cloud covered or other unusable data), and anomaly (e.g., identified as landslide despite cloud cover). A confidence level is also calculated for each cell state.  The method continuously refines cell states by analyzing time series data from successive images, reducing unknowns and anomalies to improve landslide detection accuracy. Alternating labels are considered as an indication of uncertainty, whereas cells without a clear pattern are classified as unknown. The method generates robust, time-aware landslide maps by integrating spatial classification from model outputs and cloud masks with temporal consistency checks. 

We present an overview of the developed method and demonstrate its practical application through a case study conducted in Adıyaman, Türkiye. The study focuses on landslides triggered by the February 2023 Türkiye earthquake sequence and a subsequent rainfall event in March 2023. Using various landslide detection methods (e.g., an NDVI-based approach and a deep learning model) and optical EO data with different ground resolutions (e.g., Sentinel-2, Planet SuperDove), the case study showcases the method’s ability to enhance temporal insights into landslide occurrence and progression. These results underline its potential as a valuable tool for rapid hazard monitoring and disaster response.