TL;DR
This study uses Sentinel-1 satellite data and the SBAS-InSAR technique to map ground deformation in China's Erhai region from 2015-2018, revealing significant deformation in the south driven by urbanization. They then successfully employed a back-propagation neural network to predict this deformation, achieving high accuracy and highlighting the potential of machine learning for this application.
What Problem Does This Paper Address?
Ground deformation, often caused by natural or human activities like excessive groundwater withdrawal or urban construction, poses a significant hazard. While Interferometric Synthetic Aperture Radar (InSAR) is a powerful tool for monitoring these movements, understanding the complex interplay of influencing factors and accurately predicting future deformation remains a challenge. This paper specifically addresses the need for a comprehensive analysis and predictive model for ground deformation in the Erhai region, a rapidly developing area in China, where urbanization is a growing concern. Existing prediction methods often rely on simpler linear models, which may not capture the full complexity of the deformation process.
Methods
The researchers employed the Small Baseline Subset Interferometric SAR (SBAS-InSAR) technique to analyze ground deformation. This method is an advancement over traditional InSAR, reducing errors caused by spatial and temporal decorrelation and atmospheric effects by using multiple interferograms with short baselines. They processed Sentinel-1 SAR images from both ascending and descending orbits to ensure cross-validation and accuracy.
For predicting future ground deformation, they utilized a back-propagation (BP) neural network. This is a supervised machine learning algorithm that learns from input data and their corresponding desired outputs. In this study, influencing factors like building area, water level, cumulative precipitation, and cumulative temperature were used as input parameters, with the cumulative ground deformation serving as the output. The BP network was trained and validated to predict deformation with high precision.
Datasets
- Sentinel-1 SAR Images: A total of 44 images acquired between October 2014 and December 2018, covering both ascending and descending orbits, were used for InSAR analysis.
- Landsat 8 OLI Images: Selected images from 2014 to 2018 were used to derive building area information in the southern Erhai region.
- Temperature and Precipitation Data: Acquired from the National Meteorological Information Center, these datasets were spatially interpolated to generate monthly grids.
- Water Level Data: Daily water level data from monitoring sites in the Erhai region were resampled to monthly averages.
Key Insights and Ideas
- Urbanization as a Primary Driver: The study strongly suggests that the continuous increase in building area, indicative of rapid urbanization, is the primary driver of ground deformation in the southern Erhai region.
- Seasonal vs. Long-Term Deformation: Cumulative precipitation and temperature were identified as key factors contributing to seasonal ground deformation, while building area and water level played a more significant role in long-term, cumulative deformation.
- BP Network's Predictive Power: The successful application of the back-propagation neural network demonstrates its significant potential for accurately predicting ground deformation trends, outperforming simpler linear regression models.
- Regional Disparity: A clear difference in deformation patterns was observed between the northern (stable) and southern (deforming) parts of the Erhai region, highlighting the localized impact of development.
Key Results
- Deformation Mapping: SBAS-InSAR analysis revealed that ground deformation in the Erhai region ranged from -33 to 9 mm/year between 2015 and 2018. The northern part showed minimal deformation, while the southern part exhibited significant subsidence, with some points reaching an average of -33 mm/year.
- Prediction Accuracy: The BP neural network achieved a high correlation coefficient (R) of 0.966 for predicting cumulative ground deformation. The Root Mean Square Errors (RMSE) for training, testing, and validating samples were 3.063, 1.003, and 1.119 mm, respectively, indicating high precision. The testing RMSE of 1.003 mm suggests the model can predict deformation within a millimeter of accuracy.
- Factor Correlation: The study found strong correlations (R²) between cumulative deformation and cumulative temperature (0.764), cumulative precipitation (0.746), and building area (0.431). Water level showed a weaker correlation (0.113).
Limitations and Future Work
- Limitations: The authors acknowledge that the SBAS-InSAR method may be prone to errors in mountainous areas due to slope instability and vegetation cover. While the BP model showed high precision, the paper mentions that overtraining can lead to overfitting, and the training process was stopped after 67 iterations to avoid this.
- Future Work: The study suggests that incorporating more advanced machine learning methods, such as Deep Learning (DL), could provide even greater predictive accuracy by utilizing longer time series of high-resolution data and deeper network architectures. This could offer more valuable insights for urban planning, infrastructure development, and disaster risk reduction. The authors also emphasize the need for continued monitoring and early warning systems in the southeastern Erhai region.
Summary updated