Deep Learning Recognition Method For Satellite Remote Sensing Of Complex Crop Systems

The agricultural production is inextricably linked to the well-being of the nation and its populace.China,being a large agricultural country,has a relatively large area of arable land.It is crucial to understand the agricultural production situation and adjust the planting structure to ensure the success of the "Three Rural Issues".Remote sensing imagery,with its broad coverage and easy accessibility,offers a promising solution.By analyzing the rich spectral information of target objects,it is possible to effectively identify the distribution of crops.However,the current mainstream crop identification methods rely on manual feature selection,making it challenging to identify crops with complex structures.Deep learning has powerful feature learning capabilities,which can accurately and efficiently extract complex features from images and identify complex crop systems.This research focused on single and double-season rice,which involves time complexity,and dike-pond system,which involves spatial complexity.Two algorithms,semantic segmentation and object detection,were used to investigate the potential of deep learning in identifying complex crop systems in medium and high-resolution images.The study area of Qichun County,Hubei Province was selected for the identification of single and double-season rice.Single-time semantic segmentation network and time-series network were respectively constructed for the identification based on Sentinel-2 monthly composite images from March to November 2019.For the identification of ditched rapeseed,a modified Cascade R-CNN(mCascade R-CNN)object detection network based on high-resolution satellite images was proposed and compared with Cascade R-CNN and YOLOv4.The effectiveness of mCascade R-CNN in detecting crops with spatial complexity was validated in the detection of ditched rapeseed in Qianjiang City,Hubei Province.The specific conclusions of this study are as follows:(1)A single temporal semantic segmentation network,ResNeSt＿UNet++,based on the encoder-decoder architecture was constructed.The encoder used was ResNeSt and different encoders and decoders were compared.Results showed that ResNeSt＿UNet++achieved the highest accuracy,with an F1-score of 93.68% and MIo U of 87.18%.Under the same conditions,ResNeSt outperformed ResNet and UNet++ outperformed U-Net,with increased accuracy.ResNeSt＿UNet++ also yielded smoother edge detection of paddy fields than the other models did.(2)A time series network TempCNN for direct classification of single-and doublecrop rice was constructed.TempCNN used causal convolution,which prevented future information leakage while exploiting the parallelization advantage of convolution.Compared with LSTM,SVM,and RF,TempCNN achieved the highest accuracy,with an overall accuracy of 99.36% and Kappa coefficient of 99.02%.The precision and recall rates for each class were better than those of the other three models,and the parameter size was only 2.67 MB,significantly smaller than that of LSTM(12.6MB).(3)The comparison of the above two rice recognition methods demonstrated that TempCNN excelled in crop recognition with the temporal complexity.The single-and double-crop rice extracted by TempCNN were similar to those in the Statistical Yearbook,with an additional 4640 hectares of area,including 3550 hectares of single-crop rice and1090 hectares of double-crop rice.(4)The proposed mCascade R-CNN replaced the conventional convolution kernel with DCNv2,taking into account that the direction and shape of rapeseed in remote sensing images could be irregular.This method achieved the highest accuracy in detecting dikepond systems than Cascade R-CNN and YOLOv4 did,with an AP of 80.90%,but the minimum FPS was only 4.6.The bounding boxes detected by mCascade R-CNN better fit the actual targets.Under the premise of not considering real-time performance,mCascade R-CNN was suitable for crop recognition with spatial complexity.The results indicated that the proposed method effectively identified complex crop systems in satellite remote sensing images,further improving the application of remote sensing images in crop classification and providing data support for agricultural regulation.