SAR images for benchmark dataset creation, urban polycentricity and deformation dynamics

In recent years, significant advancements in SAR sensors have resulted in a substantial increase in SAR images for urban environment applications. Effectively and massively using those remains challenging. In this thesis, I focus on applications in three aspects: benchmark dataset for training deeplearning models in land cover classification, land cover change detection for urban polycentricity analysis, and localized deformation detection and prediction.
First, a repeatable and scientific framework is proposed for creating SAR benchmark dataset for land cover classification. The framework considers three categories of SAR images: single polarimetric features, multipolarimetric features, and additional geospatial data. A hybrid radarcoding method is introduced to convert reference data into radar coordinates, aligning the radarcoded reference data with feature maps to form the SAR benchmark dataset. The attribute quality of the dataset is assessed by means of local and global attribute tables, and quantitative quality is determined by considering variance, mismatch, and uncertainty of feature maps. The dataset is stored in SpatioTemporal Asset Catalogs (STAC) format for easy sharing. The framework has been applied in Groningen, the Netherlands, using Sentinel-1 SAR images. The result demonstrates its effectiveness in generating high-quality benchmark dataset for land cover classification.
Second, a Kittler and Illingworth with Modified Model (KI-MM) method is proposed for detecting land cover changes in urban environments. This method leverages statistical information from SAR images to define a threshold that separates changed and unchanged pixels. The obtained identical threshold regarding the varied distribution of SAR images ensures the experiment is repeatable and scientific. The detected changes are linked to polycentric urban development (PUD) analysis and quantified using mean distance and mean patch area parameters. The method has been applied in Shanghai, China, using time-series Sentinel-1 and Sentinel-2 images. The result shows a successful pioneer application of SAR images in PUD analysis with the KI-MM method.
Third, a spatiotemporal concatenation method is proposed to detect wide-range deformation based upon persistent scatterers (PS). To assess the precision and accuracy, posterior variance and double difference validation have been used. The deformation results were used as input for hierarchical classification method and Getis-Ord Gi statistic to identify localized deformation in urban environments. Taking advantage of big intermediate InSAR products, such as a large number of (un)wrapped interferograms, a High-resolution network (HRNet) was trained to predict deformations. The methods have been implemented in highly developed Western Europe countries using three Sentinel-1 SAR frames. Results showed the efficiency of the concatenation method and its ability to detect and predict localized deformation.
In summary, this thesis has enriched the SAR benchmark dataset for land cover classification, strengthened analysis methods for polycentricity development, and detected and predicted localized deformation. This work thus provides a systematic approach to monitoring urban land dynamics using SAR images.