| Aquaculture has developed rapidly under the growing demand for food,huge economic interests and local policies,making important contributions to national food security and economic development,especially in the coastal area,the unique geographical advantages and suitable climatic environment provide very favorable conditions for the rapid development of the nearshore aquaculture industry.However,the rapid and disorderly development pattern of nearshore aquaculture has led to high pollution of nearby seawater and increased vulnerability of ecosystems in nearshore areas,contrary to the concept of sustainable development.Guangxi Beibu Gulf is an important sea area in the south of China,with rich fishery and aquatic resources,pond aquaculture is one of the important local aquaculture methods,facing problems such as aquaculture pollution,restricting the healthy development of local aquaculture.Therefore,accurate and rapid understanding of the spatial distribution of aquaculture ponds in the coastal area of Beibu Gulf of Guangxi is very important for the optimization of the spatial layout of aquaculture,the scientific management of natural resources and the protection of the ecological environment.The rapid development of satellite remote sensing technology has provided reliable data support for accurate and efficient extraction of nearshore aquaculture ponds.Based on multi-source remote sensing image data,this paper proposes an automatic extraction method for nearshore aquaculture ponds in Beibu Bay in Guangxi,and analyzes the temporal and spatial variation characteristics of nearshore aquaculture ponds in Beibu Bay,Guangxi.The main research contents and conclusions are as follows:(1)Using Landsat5,Landsat8 and Sentinel-2 and other multi-temporal remote sensing data,combined with Google high-definition image auxiliary data,an automatic extraction method for breeding ponds based on GEE cloud platform is proposed.Firstly,the NDWI index and Otsu algorithm were used to determine the optimal segmentation threshold to extract the aquaculture water body(including aquaculture ponds and other water types),and then,the aquaculture water body was extracted by secondary classification of the aquaculture water body by random forest algorithm,and finally the spatial distribution information of aquaculture ponds in the coastal area of Guangxi Beibu Gulf from 2000 to2022(2000,2010,2016 and 2022)was obtained by post-treatment,and the overall accuracy of extraction was above 89%.The Kappa coefficient reaches more than 0.87.Based on the extraction results of stage 4 aquaculture ponds,the area of aquaculture ponds in Beibu Bay and other regions of Guangxi in 2000,2010,2016 and 2022 was counted,and the culture center of mass in each year in the area was calculated.The results showed that the area of coastal aquaculture ponds in Guangxi Beibu Bay from 2000 to 2022 showed a substantial increase,and the area increased by 240.6km~2 in 22 years,of which2000-2010 was a period of rapid growth,and after 2010 began to transform into slow growth,Beihai City is the city with the largest area of aquaculture ponds in Beibu Bay in Guangxi,accounting for more than 60% all year round.The overall manifestation of the aquaculture center of mass in Beibu Bay of Guangxi was to migrate to the northwest,with a total of 25.1 km in 22 years.(2)Aiming at the problem that the first method is difficult to extract monomer culture ponds under the background of complex environment of remote sensing images,a semantic segmentation model of aquaculture ponds is proposed.The model is mainly based on the U-Net model,and two new structures are proposed: MS and PGC.Among them,the MS structure integrates the Inception module and the void residual module to replace the traditional convolutional layer of the U-Net model,which can enhance the model feature extraction ability and effectively solve the problems of model training gradient disappearance,and the PGC structure integrates the global context module and polarization attention mechanism to maximize the reuse of the output features of the coded part of the jump connection to enhance the context semantic information and reduce redundant information interference.In order to evaluate the segmentation performance of the improved U-Net model for breeding ponds with different resolution remote sensing images,Sentinel-2 and Planet were used as data sources,and two traditional convolutional neural network models were compared with FCN8 S and U-Net models.The experimental results show that from the evaluation index,the improved U-Net model is better than other comparison algorithms in four evaluation indicators: accuracy,recall,intersection and merge ratio,and F1 score.From the perspective of segmentation results,the improved model can overcome the interference of redundant information,reduce edge adhesion between aquaculture ponds,and better realize the segmentation of single aquaculture ponds. |
PDF Full Text Request