An efficient deep learning mechanism for cross-region generalization of landslide events

Abstract:

Land-use planning and better modeling capability needed to predict, analyze and mitigate landslide hazard would require a large-scale database of landslide events with accurate location information. Our inventory consists of Pre-event and Post-event high-resolution geo-referenced satellite images. Pixel-wise spatial distribution of landslide regions in the collected Post-event satellite images were labeled as polygons by a team of expert labelers. Pre and Post event satellite images along with their corresponding polygon annotations are used to train deep learning semantic segmentation models for automatic detection and localization of landslides. The satellite images are generally collected in batches based on the geographically distinct ecoregions. The new batch generally has a different pixel distribution thus resulting in a poor performance when pre-trained models are tested directly without retraining. This creates the need to efficiently transfer visual features for generalization across ecoregions. Considering the availability of pre-trained models trained on the existing data distribution, we propose a novel mechanism to efficiently streamline the pre-trained model's architecture such that the model "learns" the new data distribution "without forgetting" the old data distribution. This proposed method allows maximum reuse of trained parameters and usage of only the new batch of data for retraining, thus increasing computational and memory efficiency. The architecture is altered such that the newly added parameters learn the new data distribution while preserving the old model capabilities. Unlike methods of fine-tuning and feature extraction, our novel architecture is trained end-to-end without freezing the old trained parameters. We observe that our method outperforms the traditional methods of domain adaptation and transfer learning, fine tuning, joint training and feature extraction, in terms of training and testing efficiencies, memory requirements and eliminates the need for the data from previous ecoregions. Qualitative and quantitative results further support our claim. Our proposed method can be used iteratively as every new batch of data flows in temporally.