Reservoir water and rainfall, leading to fluctuations groundwater levels, are the main triggering factors that induce landslides in the Three Gorges Reservoir area. This study investigates the response mechanism of landslide deformation under reservoir water and rainfall variations through long-time on-site observations. To address the non-stationary characteristics of the time-series records, joint time-frequency analysis (JTFA) is first introduced into our landslide prediction model. This model employs optimal variational mode decomposition (VMD) to obtain specific signal components with clear physical meaning, such as trend component and periodic components. Then, multi-scale response analysis between the displacement and external factors three wavelet methods was conducted. The analysis results show a 1 year primary cycle of the time series associated with the landslide evolution. The reservoir water level and rainfall show anti-phase fluctuations. The periodic displacement correlates significantly with rainfall, lagging by about two months. The reservoir water is anti-phase with the landslide displacement, preceding it by approximately three months (−51 ± 8° phase difference). For landslide displacement prediction, the gated recurrent units (GRU) neural network model is integrated into the deep learning forecasting architecture. The model takes into account the correlation and hysteresis effect of input variables. Through six experiments, we investigate the effect of data volume on model predictions to determine the optimal model. The results demonstrate that our proposed model ensures high performance in landslide prediction. Moreover, a comparison with six other intelligent algorithms shows the advantages of our model in terms of time-effectiveness and long-sequence forecasting.

Healthy coasts have a high ecological service value. However, many coastal areas are faced with oil spill risks. The Synthetic Aperture Radar (SAR) remote sensing technique has become an effective tool for monitoring the oil spill risk in coastal areas. In this study, taking Jiaozhou Bay in China as the study area, an innovative oil spill monitoring framework was established based on Polarimetric SAR (PolSAR) images and deep learning theory. Specifically, a DeepLabv3+-based semantic segmentation model was trained using 35 Sentinel-1 satellite images of oil films on the sea surface from maritime sectors in different regions all over the world, which not only considered the information from the PolSAR images but also meteorological conditions; then, the well-trained framework was deployed to identify the oil films in the Sentinel-1 images of Jiaozhou Bay from 2017 to 2019. The experimental results show that the detection accuracies of the proposed oil spill detection model were higher than 0.95. It was found that the oil films in Jiaozhou Bay were mainly concentrated in the vicinity of the waterways and coastal port terminals, that the occurrence frequency of oil spills in Jiaozhou Bay decreased from 2017 to 2019, and that more than 80 percent of the oil spill events occurred at night, mainly coming from the illegal discharge of waste oil from ships. These data indicate that, in the future, the PolSAR technique will play a more important role in oil spill monitoring for Jiaozhou Bay due to its capability to capture images at night.