Enhancing Interpretability in Landslide Detection: A U-Net Based Framework with Uncertainty-Aware Classification Hina Tachikawa1*, Yuki Mizuno1, Naoyoshi Hirade1, Kenlo Nishida Nasahara2 and Taro Uchida3
1 Graduate student, Graduate School of Science and Technology, University of Tsukuba, Japan
2 Associate Professor, Institute of Life and Environmental Sciences, University of Tsukuba, Japan
3 Professor, Institute of Life and Environmental Sciences, University of Tsukuba, Japan
htl21.tachikawa[at]gmail.com
Abstract
Automatic landslide mapping plays a critical role in disaster response and risk assessment. However, deep learning models are often prone to misclassifications, such as overlooked landslides and non landslide areas wrongly identified as landslides, which can hinder accurate interpretation and decision making. Conventional approaches typically provide only landslide or non landslide classes without indicating the chance of misclassification. Consequently, smaller landslide areas are more likely to be overlooked in comparison to large ones, and false readings are more likely to occur at the edges of landslide. To address this limitation, we propose a four category classification scheme that improves interpretability by distinguishing between reliable results and areas with high misclassification possibilities. We applied Monte Carlo dropout to a U-Net based model trained on three landslide events in Japan when testing on a separate area. Using 20 stochastic forward passes, we computed the mean and standard deviation of landslide probabilities. A low standard deviation indicates a reliable result, while a high standard deviation suggests a high chances of misclassification. Based on these values, each pixel was classified into one of four categories, Landslide, Non landslide, Potential false negative, and Potential false positive. The proposed method successfully visualized small landslide and landslide edges as regions with high probability of misclassification. This enhanced the interpretability of the output and demonstrated the potential in improving the trustworthiness in landslide detection. The proposed four class classification scheme offers a novel way to assess and communicate model confidence, addressing a key limitation of conventional binary classification approaches. For instance, considering potential false negative class let us examine the safety aspect. Alternatively, focusing exclusively on reliable classes let us consider minimum required response.
Keywords: landslide detection, deep learning, U-Net, Monte Carlo dropout, satellite remote sensing
