Uncertainty evaluation and proposals for performance improvement methodologies in image generation of landslide mass by deep learning models

• Published in: Journal of the Japan society of photogrammetry and remote sensing Vol. 61; no. 6; pp. 368 - 386
• Main Authors: NAKAMURA, Yoshio, TAKEUCHI, Yutaro, FURUKI, Hirokazu, YOSHIDA, Kazuya, UTSUKI, Shinji, YAMAMOTO, Yoshiyuki
• Format: Journal Article
• Language: English; Japanese
• Published: Tokyo Japan Society of Photogrammetry and Remote Sensing 2022; Japan Science and Technology Agency
• Subjects: Deep learning; Image processing; Landslides; Proposals; Recall; Uncertainty
• ISSN: 0285-5844; 1883-9061
• DOI: 10.4287/jsprs.61.368

This paper shows uncertainty evaluation and proposals for performance improvement methodologies in image generation of landslide mass by deep learning models. Model ensembles and Monte Carlo Dropout (MCD) were used to evaluate uncertainty. Three methodologies were examined as performance improvement methodologies. The methodologies were slide processing, recall/precision emphasized models, and transfer learning with an inherent factor of landslide. The recall/precision emphasized models were developed by the improved loss function. The result showed that MCD could not be an alternative to model ensembles. In performance improvement methodologies, the transfer learning with geology distribution scored at 80% of precision. The recall/precision emphasized models inferred the distribution of landslide mass adequately. The effectiveness of the slide processing was found to be dependent on the performance of the trained model.