Research On Land Cover Change Detection Method Based On Quad-pol Radar Remote Sensing Data

Land use and land cover(LULC)change detection uses systematic algorithms to compare and analyze remote sensing images in different periods to identify land surface changes in specific area.China's urbanization and rural infrastructure construction are advancing steadily,and land cover is changing rapidly.The demand for land resources and the consequential impact on land changes caused by human activities are increasing with the population explosion.Optical imaging is vulnerable to meteorological interference,making it difficult to obtain high-quality optical images continuously,and to support land change detection robustly.Satellite equipped with Synthetic aperture radar(SAR)sensors can work in all-day and all-weather condition,thus is an important alternative in foggy and cloudy area.The traditional pixel-based algorithms bring out significant salt and pepper noise in change detection results,the coherent speckle noise of SAR image further reduces the detection accuracy.In recent years,deep learning algorithms achieved good results in natural image classification and pattern recognition.Deep learning can automatically extract remote sensing image features,mine complex parameter relationship,and have good fault-tolerant capacity.However,SAR data acquisition cost is high,and SAR visual interpretation is difficult.Therefore,SAR-based change detection using deep learning algorithms need to be further studied.Based on the fully polarized Radarsat-2 data,this paper takes Meishan area in Sichuan Province as an example to carry out land cover change detection research.The main work and results are as follows:(1)This thesis proposed a change detection method based on object-oriented recognition using optical information to guide SAR difference images segmentation.The SAR difference image is generated from the pre-and post-change SAR images,and the post-change Sentinel-2 optical image is used to guide the object-oriented segmentation process of the SAR difference image,to extract multi-dimensional features at object-level,and to optimize the feature space.Land cover change detection is achieved combining object-oriented recognition with random forest algorithm.Compared with the existing research,the accuracy of this method is markedly improved(accuracy=92.90%,recall =96.61%,F1-measure=96.07%,commission=4.47% and omission=3.39%),which can effectively reduce salt and pepper noise.(2)This thesis evaluated and analyzed the feasibility and performance of applying deep learning model to land cover change detection based on full polarimetric SAR data.The deep separable convolution of Mobile Net model is used to replace the traditional convolution for feature extraction.Combined with the basic framework of U-Net model,a Mobile UNet model is proposed to adapt to small sample training.Different models are tested based on SAR datasets,and the results show that deep learning modes are also suitable for change detection applications using SAR data,and can obtained effective performance.Mobile UNet model has the highest accuracy(accuracy=95.11%,recall=96.15%,F1 measure=97.15%,commission=1.79%,and omission=3.85%)for regions with remarkable and complex land changes.(3)This thesis performed noise sensitivity analysis for the above two methods.Using SAR images speckle-filtered by Refined Lee algorithm with different spatial windows,and un-filtered SAR images,different datasets are built,based on which land cover change detection is carried out to explore the sensitivity of the two methods to radar noise.The results show that both methods are significantly affected by noise,and the change detection result of the deep semantic segmentation method shows certain regularity due to the balance between signal information and noise under different filtering windows.The preservation of image details and the removal of noise can be well balanced,resulting in the best change detection effect when the filtering window is 7×7.