HADeenNet: A hierarchical-attention multi-scale deconvolution network for landslide detection

Abstract

Efficient landslide mapping from high spatial resolution images is important in many practical applications, such as emergency response. Numerous studies and methods have been published on this subject; however, these methods are difficult to apply in the real world because they are mainly based on remotely sensed landslides from a single sensor with a specific spatial resolution. Additionally, models built within deep learning frameworks tend to adopt similar encoder-decoder network structures, wherein many landslide features are easily filtered out by continuous convolutions. In this paper, we propose a hierarchical deconvolution network to detect landslides. The model enlarges input feature maps by a deconvolution operation and convolutes the enlarged feature maps to learn to detect landslides. Moreover, the hierarchical structure enables the proposed network to better synthesize landslide features at a higher spatial resolution. An attention module is also proposed to enhance multi-scale landslide features. Our model is trained on four large earthquake-triggered landslide areas and one publicly released landslide dataset, where each dataset consists of hundreds to thousands of landslides. To mimic practical applications, the trained model is evaluated over three areas that recently experienced landslides. The performance of our proposed model is compared with six widely used frameworks, and it achieved a 21% higher F1-measure and at least 10% higher IOU using each of the evaluation landslide datasets. Additionally, the effectiveness of our model in maintaining landslide features, especially small landslides, is verified through comparisons with other commonly used frameworks, which demonstrate a strong potential use for practical cases.