| The rapid extraction of information on the distribution range,quantity and area of aquaculture zone from high-resolution remote sensing images is of great significance for site planning,yield estimation and ecological environmental protection of aquaculture zone.With the development of high-definition satellite technology in China,it has become easier and easier to obtain high-resolution remote sensing satellite images.The traditional method of extracting aquaculture zone is time-consuming and difficult to guarantee accuracy.This paper introduces the FCN network to the extraction of farming areas from high-resolution remote sensing images,and investigates three aspects: the aquaculture zone sample library based on high-resolution images,the aquaculture zone extraction model based on FCN network,and the application of extraction of aquaculture zone from pits and net pens based on high-resolution images.(1)The sample library of farming area based on high resolution image.In this paper,the sample libraries for farming area extraction were constructed by Labelme annotation software and high resolution images,including 1623 sets of 0.8m pit culture sample library,1623 sets of 2m pit culture sample library,2784 sets of 0.8m net tank culture sample library and 2784 sets of 2m net tank culture sample library,which provide a foundation for the subsequent training tests of the model and exploring the influence of different resolution of images and farming types on the final The results of the extraction were based on the following model training and testing,as well as exploring the influence of different image resolutions and culture types on the final extraction results.(2)FCN network based farming area extraction model.In this paper,the trained FCN network is tested for farming area extraction on the sample library test set.The farming extraction method based on deep learning FCN network is able to perform targeted extraction of farming areas and has high extraction accuracy.In this study,the accuracy was evaluated and analyzed by accuracy,recall,F-value and kappa coefficient.The F-value of 0.8m pit pond culture extraction model was 0.96;the F-value of 2m pit pond culture extraction model was0.93;the F-value of 0.8m net tank culture accuracy was 0.94;2m net tank culture extraction model;the kappa coefficients of the four culture area extraction models were 0.93,0.90,0.93,0.82,and almost identical in terms of consistency.(3)Application of extraction of pit pond culture and net box culture based on high-resolution images.In this paper,Huzhou City,Zhejiang Province and Minjiang River(Gutian section)were selected as the test area,and the extraction of information based on high-definition imagery for pond farming and net-pen farming was carried out respectively.The extracted results were vectorised,co-ordinate aligned and co-ordinate transformed to calculate the farming area of the study area and compared with the real farming area at the corresponding time to verify the accuracy of the extraction results.This paper constructs a sample library of high-resolution images of aquaculture zone and establishes a deep learning-based farming area extraction method that can obtain data on the distribution,number and area of aquaculture zone more quickly and accurately,which can provide a new method for the extraction of aquaculture zone. |
