| Tidal flat refers to the tidal flooding zone between the high and low tides of the coastal tide,which is made of deposit mud and sand.Tidal flat are important wildlife habitats for a large number of waterfowl,migratory birds,crabs and fish,and also are important material foundations and reserve resources for the development of coastal cities.However,tidal flats are affected by coastal development,sea level rise,coastal erosion,increase and decrease in river sediment flow,and the settlement and compaction of coastal sediments.By observing the remote sensing images,we can notice the change of tidal flats and protect and restore the coastal ecosystems.Due to the action of gravitation,the tidal flats are sometimes submerged by water and sometimes exposed to the water surface.In the process of dynamic changes,the accurate range of the tidal flats is difficult to define.Therefore,it is necessary to use multi-temporal remote sensing images to observe the tidal flats.Traditional observation methods perform supervised learning on the multi-period bands and remote sensing coefficients on pixel level,such as the random forest method,which is not only time-consuming but also unable to integrate the surrounding pixel information.Deep learning methods are currently widely used in change detection tasks,however,it hard to deal with the multiple input images and lack of labels.This research is based on the project of "Middle and Lower Reservoirization Driven by the Yangtze River-Unbalanced Conversion of Estuary Landforms,Environment and Ecology".Multi-temporal remote sensing satellite images are synthesized into the dual-phase images,highest and lowest tide level images,which are used for change detection.The specific work is as follows:(1)We proposed a novel framework based on Adapt Seg Net model that combined with remote sensing derived indices to segment remote sensing images by water and land,the model was trained on publicly labeled GID dataset and the Landsat dataset made by ourself.This task is to alleviate the data distribution discrepancy of feature between the two datasets meanwhile keeps the detail information in target dataset and solves the lack of labeling in the target datasets.The NDWI water body value was used to strengthen the segmentation details and improve the accuracy.Our method achieved 90.0% mean intersection-over-union.experiments showed that the result is better than the baseline model and other contrast model.Besides,this model provided a robust backbone network for the subsequent siamese network structure.(2)We proposed a model based on the siamese network to detect changes on tidal flat.First,the multi-temporal remote sensing image were converted into a dual-phase image(the highest tide level and the lowest tide level image)through the MSIC method as the input of the model.We integrated the feature pyramid module to improve the model effect,and used the same structure and parameters trained from sea-land segmentation for transfer learning.(3)We also designed a structure to segment the tidal flat on panoramic Landsat images and analyzed the change of tidal flat since 1984.Specifically,we first synthesize the annual image on the GEE(Google Earth Engine)platform,then crop the super-large pixel remote sensing image,put them into the trained model and merge the result images in post-processing operations,finally the area is calculated by the number of pixels obtained by the segmentation map.In this paper,we sampled and selected tidal flats in the eastern coastal area of China,produced 4000 tidal flat change detection data sets and conducted detailed experiments.Finally,we achieved 79.0% of the dice coefficient,which were better than the contrast method.The ablation experiment proved the effectiveness of our improvement and raised 2.5% of dice coefficient... |
PDF Full Text Request