Automatic detection of land cover changes using multi-temporal polarimetric SAR imagery

Abstract

Dramatic land-cover changes have occurred in a broad range of spatial and temporal scales over the last decades. Satellite remote sensing, which can observe the earth's surface in a consistent manner, has been playing an important role in monitoring and evaluating land-cover changes. Meanwhile, optical remote sensing, a common approach to acquiring land-cover information, is limited by weather conditions and thus is greatly constrained in areas with frequent cloud cover and rainfall. Recent advances in polarimetric SAR (PolSAR) provide a promising means to extract timely information of land-cover changes regardless of weather conditions.

SAR satellite can pass through an area from different orbits, namely ascending orbit and descending orbit. The PolSAR images from the same orbit will have similar backscattering even with different incident angles. But if images are acquired from different orbits, the backscattering will vary greatly, which causes many difficulties to land cover change detection. The proposed algorithms in this study can perform land cover change detection in three situations: 1) repeat-pass images (image from the same orbit and with same incident angle, 2) images from the same orbit but with different incident angle, and 3) images from different orbits. Using images from different orbits will largely reduce the monitoring interval which is important in the surveillance of natural disasters.

The present study proposes 1) a sub-pixel automatic registration technique, 2) an automatic change detection technique and 3) an iterative framework to process a time series of PolSAR images that can be applied to the PolSAR images from different orbits. Firstly, automatic registration is crucial to the change detection task because a small positional error will largely degrade the accuracy of change detection. The automatic registration technique is based on the multi-scale Harris corner detector. To improve the efficiency and robustness, the orientation angle differencing method is proposed to reject outliers. This differencing method has been proved effective even in the experiment of using PolSAR images from different orbits when less than 5% of the feature point matches are correct. Secondly, the change detection technique can automatically detect land-cover conversions and classify the newly input image. Hierarchical segmentation has been applied in the change detection which generates objects within the constraint of the previous classification map. Multivariate kernel density estimation is applied to classify newly input PolSAR image. The experiments show that the proposed change detection technique can mitigate the effect of polarimetric orientation shift when the PolSAR images are from different orbits, and it can achieve high accuracy even when complex local deformation is appeared. Lastly, the iterative framework, which integrates the automatic registration and automatic change detection techniques, is proposed to process a time series of PolSAR images. In the iterative process, no obvious decrease of classification accuracy is observed. Therefore, the proposed framework provides a potential treatment to derive land-cover dynamics from a time series of PolSAR images from different orbits.