Land Use Classification And Change Detection In Coastal Areas Based On Timeseries SAR Images

Land use change is of great significance for analyzing climate and ecological environment changes,whereas obtaining the land use information is the basis for achieving this goal.The development of multi-source remote sensing platforms has facilitated the rapid and effective acquisition of the land use patterns.However,China’s coastal cities are susceptible to typhoon and other severe weather conditions.Cloud,rain and fog weather lead to poor optical image quality.Furthermore,the repeat cycle of remote sensing platforms is generally longer,resulting in a reduction in the number of remote sensing images that can be used in coastal cities,which poses difficulties in obtaining the land use information and analyzing the land use changes.In addition,coastal cities are susceptible to extreme flood disasters due to the global climate change and Typhoon-dominated climate in coastal zones.Flood disasters inundated the infrastructure,large areas of farmland and houses in coastal areas,even caused the changes in human settlements,which led to changes in the spatial pattern of land use.Moreover,flood disasters cause changes in the structure of biological populations,which may disrupt the original ecological balance and deteriorate the ecosystem,even cause other environmental problems.And flood disasters will also bring other disasters,such as water pollution,obstructed transportation,and even cause outbreaks and epidemics of infectious diseases.Considering the limited application of optical remote sensing images in coastal areas,the synthetic aperture radar(SAR)images are not affected by the weather,and with high spatial resolution and short revisit period.Therefore,it is urgent to integrate multisource images such as optical and SAR images to detect the change of land use pattern in coastal areas and monitor and extract the coastal submerged area in real time.Considering the above proposed issues,this paper takes the coastal city-Taizhou as the study area,which is located in Zhejiang Province.First,the SAR image is processed by filtering,geocoding,and radiation calibration.The optical image is processed by radiation correction and atmospheric correction.The optical and SAR images are coregistered to achieve the matching of corresponding pixel spatial positions in the two images.Secondly,the standard typical image fusion methods were applied to fuse the optical and SAR images,including HSV transform,principal component transform and Gram-Schmidt transform.Based on the optimal fusion image,three conventional supervised classification methods were used to interpret the land use pattern,including statistical-based maximum likelihood classifier,machine learning-based support vector machine and neural network classifier.Then we calculate the transfer matrix of the obtained land use pattern to analyze the land use change.Finally,based on the pre-and post-disaster SAR images,the difference method and the threshold method are used to extract the flooded area,and based on the land use classification results,the flood inundation losses are calculated.The main conclusions of this paper are as follows:(1)Three pixel-level-based fusion algorithms show that the Gram-Schmidt method is optimal for the fusion of optical and SAR images.The visual inspection shows that the fusion of HSV and Gram-Schmidt transformation results in large differences in the color of the fusion images and the texture information of the SAR image is preserved.However,HSV only fuses the three bands,and the quantitative assessment shows that GramSchmidt fusion image generated the largest information entropy and the contrast of the result is higher.The comparison of the three fusion results shows that Gram-Schmidt can better increase the image information and the separability of the features.(2)The accuracy generated by the three classification results shows that the maximum likelihood method produces the highest overall accuracy and Kappa coefficient,which are 92% and 89%,respectively.The classification accuracy of the neural network is worse,which is 6% and 8% lower than the accuracy produced by the maximum likelihood method,respectively.But the support vector machine has the lowest accuracy with the overall accuracy of 83% and Kappa coefficient of 77%.Accuracy evaluation shows that applying the maximum likelihood method to classify the fused image can obtain more accurate land use patterns.(3)Analysis of the land use change in Taizhou from 2015 to 2019 shows that its land use pattern has remained basically stable,and no major changes have occurred in the pattern of the main land use types.Compared with the land use area in 2015,the area of agricultural land decreased by approximately 16.85%,the forest land increased by 280 square kilometers,while the area of construction land exceeded 1,000 square kilometers.Little change in water area,and the area of unused land decreased by approximately 26 square kilometers,accounting for about 46.41% of its area in 2015.(4)Based on the pre-and post-disaster SAR images of Linhai City,the flood submerged area is detected and extracted.The flood inundation area caused by typhoon "Lekima" reached 135.4 square kilometers,and the main type of the inundated land was agricultural land,and the submerged area reached 97.26 square kilometers.During the detection and extraction of the flooded area using difference and threshold methods,the initial extraction area was 148.5 square kilometers,but when the land use disaster loss statistics were performed,the water area was 13.1 square kilometers.The flood was mixed with sediment and the timber that flow into the water,causing the changes in water quality and the water was extracted as flood.Therefore,the actual submerged area of this flood disaster is 135.4 square kilometers.(5)Through the study in this paper,it is found that the fusion of optical and SAR images improves the image information and the separability of ground features,which provides a better data foundation for the acquisition of land use information.The statistics method generated a higher speed and accuracy than the machine learning methods,but it needs further study and validation.Different temporal SAR images can be used to detect the spatial distribution of the flooded areas,and the difference and threshold method can be applied to quickly detect and extract the flood submerged area from SAR images.