A multi-modal deep learning framework with GAN-based fusion for enhanced landslide detection

Abstract

The proposed work presents a hybrid deep learning framework that integrates four pre-trained Convolutional Neural Networks that include VGG16, DenseNet201, ResNet50 and InceptionV3. The pre-trained CNNs are combined with a GAN-based adversarial refinement module for accurate landslide detection and segmentation. Unlike traditional single-CNN or ensemble models, the proposed model performs multi-backbone feature fusion to accurately capture global level terrain context and fine-grained spatial details. The GAN component sharpens boundaries and suppresses noisy predictions through discriminator-guided refinement. The proposed system generates GIS-ready probability maps with confidence layers. They are also optimized for low-latency inference, making it suitable for rapid post-disaster decision support. The proposed work is evaluated on three benchmark datasets - CAS Landslide (high-resolution GF-2/UAV imagery), MS2LandsNet (medium-resolution Sentinel-2) and GDCLD (coseismic landslides). The proposed framework achieves F1-scores of 97.24%, 93.70% and 94.75% across the three datasets. These results correspond to improvements of 1.4 to 2.9% over fusion baselines and 4–7% over single-CNN models such as VGG16, DenseNet201,ResNet50 andInceptionV3. The results highlight consistent IoU gains and improved boundary delineation. The cross-dataset experiments further demonstrate strong generalization across varying resolutions, terrain types and triggering mechanisms. To our knowledge, this is the first landslide segmentation study to combine multi-backbone feature fusion with adversarial mask refinement in an operational monitoring context. The results confirm that the proposed framework delivers high accuracy, scalability and deployment readiness making.