<p>Early potential landslide hazards are highly concealed, with sparse and subtle feature information. Current deep learning methods for potential landslide hazard identification struggle to model multi-source data’s nonlinear interactions and integrate multi-scale features. This leads to difficulties in identifying small potential landslide hazards in complex areas, resulting in high missed detection rates. To address this, we propose an adaptive potential landslide hazard identification method based on a KAN-UNet model integrated with multi-source remote sensing data. First, we constructed a multi-source potential landslide hazard dataset by integrating Sentinel-1A, Sentinel-2A, and DEM data, incorporating multi-dimensional features from optical bands, spectral indices, topographic features, and deformation data. Next, we developed a potential landslide hazard identification model capable of dynamically modeling complex nonlinear interactions between multi-scale features using B-spline basis functions. This was achieved by replacing U-Net’s traditional convolutions with OptimizedKANConv2D and introducing a TokenizedKANBlock in the deeper encoder to enhance feature extraction. Finally, by combining a MultiScaleFusion multi-scale gated fusion mechanism with elastic deformation and CutMix hybrid augmentation strategies, we improved the model’s accuracy and robustness for small-target potential landslide hazards in complex areas. We evaluated our method in Lanzhou City. Experimental results show that, compared to using only optical bands, integrating multi-source data improved IoU and F1-score by 4.62% and 3.37%, respectively. Furthermore, our KAN-UNet model achieved the highest IoU, precision, recall, and F1-score among other improved U-Net models, with IoU increasing from 67.95% to 89.80% and the missed detection rate dropping from 21.21% to 6.50%. This study offers a new approach for intelligent geological hazard monitoring in complex geomorphic regions.</p>