Sie befinden Sich nicht im Netzwerk der Universität Paderborn. Der Zugriff auf elektronische Ressourcen ist gegebenenfalls nur via VPN oder Shibboleth (DFN-AAI) möglich. mehr Informationen...
Landslide detection from bitemporal satellite imagery using attention-based deep neural networks
Ist Teil von
Landslides, 2022-10, Vol.19 (10), p.2459-2471
Ort / Verlag
Berlin/Heidelberg: Springer Berlin Heidelberg
Erscheinungsjahr
2022
Link zum Volltext
Quelle
SpringerLink
Beschreibungen/Notizen
Torrential rainfall predisposes hills to catastrophic landslides resulting in serious damage to life and property. Landslide inventory maps are therefore essential for rapid response and developing disaster mitigation strategies. Manual mapping techniques are laborious and time-consuming, and thus not ideal in rapid response situations. Automated landslide mapping from optical satellite imagery using deep neural networks (DNNs) is becoming popular. However, distinguishing landslides from other changed objects in optical imagery using backbone DNNs alone is difficult. Attention modules have been introduced recently into the architecture of DNNs to address this problem by improving the discriminative ability of DNNs and suppressing noisy backgrounds. This study compares two state-of-the-art attention-boosted deep Siamese neural networks in mapping rainfall-induced landslides in the mountainous Himalayan region of Nepal using Planetscope (PS) satellite imagery. Our findings confirm that attention networks improve the performance of DNNs as they can extract more discriminative features. The Siamese Nested U-Net (SNUNet) produced the best and most coherent landslide inventory map among the methods in the test area, achieving an F1-score of 0.73, which is comparable to other similar studies. Our findings demonstrate a prospect for application of the attention-based DNNs in rapid landslide mapping and disaster mitigation not only for rainfall-triggered landslides but also for earthquake-triggered landslides.