Research On Fast Identification And Extraction Method Of Rice Based On Remote Sensing Image

Automatic identification and extraction of rice planting area and monitoring of planting area are of great significance to the grain yield estimation in China.Remote sensing technology has been widely used in the identification and extraction of rice planting area,the traditional method of extracting rice is time-consuming and laborious,and its shallow structure is difficult to get better classification results,while the deep learning model has gradually become an important method in the field of precision agriculture,which can automatically learn the feature information in the image and then accurately classify it.In this paper,based on summarizing the relevant research results,in order to quickly identify and accurately extract the rice planting areas in Tianjin,an in-depth research on crop classification methods was carried out,and the specific research work and results achieved are as follows:(1)In this study,Sentinel-2 images were used as the data source,and Tianjin City was selected as the experimental area,with rice as the research object.By analyzing the characteristic changes of rice spectral curves such as Normalized Difference Vegetation Index and Land Surface Water Index,calculating the average value of rice sample image elements of different indexes in each month,we derived the growth change rule of rice,and then combined with the rice phenology of Tianjin City,we determined the optimal time period for rice extraction from 2019 to 2022,and downloaded the images using the Google Earth Engine platform.At the same time,this paper provides comprehensive and systematic basic data guarantee for the research work by collecting rice sample data in Tianjin City in the field.(2)Object-oriented remote sensing image based on the rapid identification method of rice research.Aiming at the problem of time-consuming recognition of rice by deep learning methods,in the case of existing historical data of plot vector boundary,a method is proposed to process the input data of the deep learning model,based on random sampling strategy and image reconstruction method for any irregularly shaped recognition target,automatic normalized sampling processing,and visual attention network model to achieve rapid recognition of rice.The experimental results show that the method greatly reduces the amount of computation and improves the computational efficiency while taking into account the recognition accuracy.(3)Research on rice extraction from remote sensing images based on U-Net(WTU-Net)network with data set optimization and improvement.Aiming at the problem of lack of training samples and time-consuming and laborious sample production in the field of agriculture,and also in order to realize the extraction of high-precision rice planting areas in a large area,this paper proposes to adopt the Random Forest algorithm to automate the production of deep learning datasets,and to carry out the extraction of rice planting areas in Tianjin based on the optimized and improved U-Net model structure of WTU-Net network,and to combine the extraction results with the U-Net network,DeepLabV3+ network extraction results were compared and analyzed,and the results showed that the WTU-Net network rice extraction had the highest total accuracy,reaching 94.42%.The WTU-Net network was then utilized to extract rice in Tianjin from 2019 to 2022 and count its planted area,and compared with the statistical results of the National Bureau of Statistics Tianjin Municipal Survey Team,the area error of the four years was about 10%.The feasibility of WTU-Net network application in large-scale rice mapping is illustrated,and the extraction accuracy is high.