Extraction Of Rapeseed Planting Area Based On Sentinel-2 Data And Deep Learning Approach

Rapeseed is an important crop and the first oil crop in China.Obtaining prompt and accurate acreage of rapeseed is an important basis for crop growth monitoring and yield estimation,which is of great significance for agricultural production management and national oil supply security.In the early stage,the methods of obtaining crop planting area were mainly field survey and step by step statistical reporting.The current situation of the data is not high,the investigation workload is heavy and the cost is high,and there are various drawbacks.In recent years,Sentinel-2 data has provided good data support for rapid acquisition of a large range of rapeseed planting area due to its rich spectral information,wide coverage and easy access.In addition,the application of deep learning technology in semantic segmentation also provides classification algorithm ideas for precision agriculture.Based on this,Chongming Island of Shanghai was selected as the study area to obtain Sentinel-2 images of rapeseed in full bloom.The rapeseed feature data set was constructed and the feature optimization was carried out.Based on U-Net model,an extraction model of rapeseed planting area oriented to medium and high resolution was constructed,and compared with three traditional remote sensing classification methods.Combining feature selection and classification methods,a feature fusion deep learning method for rapeseed planting area extraction was proposed to provide an accurate,feasible and intelligent interpretation method for rapeseed planting area extraction.The main research work is as follows:(1)The rapeseed feature data set was constructed and the feature optimization was carried out.In this paper,vegetation index and texture features were extracted from Sentinel-2 remote sensing data to construct rapeseed feature data set,and a feature optimization method combining correlation coefficient,feature importance ranking and backward elimination feature selection was proposed.Finally,12 feature variables were selected to form the optimal feature data set.The results of feature selection comparison experiments using random forest classification algorithm showed that the addition of feature data could improve the extraction accuracy of rapeseed planting area.Among them,the classification result based on the preferred feature data set was the best,with an overall accuracy of 90.40% and Kappa coefficient of 0.8779.Compared with the classification result of the original data,the overall accuracy is improved by 4.34%,and the Kappa coefficient is improved by 0.06.The results demonstrate the effectiveness of the feature selection method.(2)Determine the best classification method for extraction of rapeseed planting area.Based on Sentinel-2 remote sensing image,four classification methods,Random Forest(RF),Support Vector Machine(SVM),Object-based(OB)classification and deep learning model(U-Net),were compared to determine the best model of rape planting area extraction.The experimental results show that the extraction accuracy of rape planting area based on U-Net model was the best,with the overall accuracy reaching 97.98% and Kappa coefficient reaching 0.8950.The experimental results show that the deep learning algorithm is feasible and accurate in the study of rape planting area extraction,which can provide an effective intelligent interpretation method for the extraction of rapeseed planting area in broken areas of South China in the future.(3)Application of deep learning fusion features in rapeseed planting area extraction.Combined with the optimal classification model and the optimal feature data,the cultivation area of rapeseed in Chongming Island of Shanghai was extracted from2019 to 2021,and its spatio-temporal variation was analyzed.The results show that the overall accuracy of rapeseed acreage extraction from 2019 to 2021 reaches 97.88% on average,and the Kappa coefficient reaches 0.9345 on average.From 2019 to 2023,the planting area of rape on Chongming Island was 4695.02 ha,5334.8 ha,2284.1 ha,1159.19 ha and 1001.46 ha respectively,mainly concentrated in Xinhai Town,Gangxi Town,Dongping Town and Chongming Dongtan,and only 4,693.5 mu of rape has been steadily planted in the past three years.The results showed that the method could provide methods and ideas for the accurate identification and change analysis of rapeseed planting area in Chongming Island,a world-class ecological island,under the cultivated land fine utilization mode of crop rotation and fallow and crop structure adjustment.The development of this study can provide certain methods and ideas for the extraction of rapeseed planting area in the regions with centralized layout and high degree of fragmentation of cultivated land in the south,and provide scientific basis for the adjustment of local agricultural planting structure and the high-quality development of modern green agriculture.