时序InSAR滑坡形变监测与预测的N-BEATS深度学习法——以新铺滑坡为例

Abstract

滑坡通常发生突然, 破坏力巨大, 经常造成重大生命安全事故和财产损失。高可靠性、高精度及具有抗差性能的滑坡形变监测预测手段和方法对于国家防灾减灾需求具有切实意义。InSAR技术是一种能够全天时和全天候观测获取高空间分辨率和宽覆盖率影像, 高灵敏性捕捉时空维动态变化的监测手段, 然而目前应用InSAR时序影像对滑坡区进行滑坡预测的工作仅是凤毛麟角。基于时序InSAR观测结果, 本文提出了一种能够有效解决中短期滑坡预测问题的深度学习滑坡预测方法。在三峡新铺滑坡区应用N-BEATS网络模型和Sentinel-1 SAR数据进行形变预测, 以均方根误差1.1mm的预测精度完成了滑坡预测工作, 并对预测结果进行了数据结构影响的规律性分析、传统方法效果对比、抗差性评估及置信区间估计等多方位的剖析, 结果显示出了其高精度、高可靠性及具有一定抗差能力的突出优势。

Alternate abstract:

Landslides usually occur suddenly and cause great damage, often causing serious life safety accidents and property losses. The monitoring and prediction methods of landslide deformation with high reliability, high precision and anti-difference performance are of practical significance to the needs of national disaster prevention and mitigation. Interferometric synthetic aperture radar(InSAR) technology is a monitoring method capable of all-day and all-weather observation, obtaining images with high spatial resolution and wide coverage, and capturing dynamic changes of spatio-temporal dimensions with high sensitivity. However, at present, the landslide prediction based on InSAR time series image is very rare. This paper presents a landslide prediction method based on deep learning, which can effectively solve the problem of medium- and short-term landslide prediction by exploiting multi-temporal InSAR observations. Neural basis expansion analysis (N-BEATS) network model was used to predict the landslide in the Xinpu area, the Three Gorges. The landslide prediction was completed with an accuracy (root mean square error) of 1.1 mm. The results are analyzed by the regularity of data structure, comparison to traditional methods, evaluation of the tolerance and estimation of the confidence interval. The results show that the proposed prediction method has outstanding advantages of high precision, high reliability and certain robust estimation ability.