Paddy Rice Mapping With Deep Semantic Segmentation Network Based On Landsat Data

Rice is the main food crop in China.Mapping rice planting areas is the basis of rice yield estimation and growth monitoring.It is also of great significance for optimizing the agricultural planting structure,promoting agricultural sustainable development and global ecological environmental change research.Compared with traditional classification methods,deep learning technology can effectively improve crop mapping accuracy.Thus,using deep semantic segmentation networks for crop mapping is currently a hot and frontier topic in crop mapping research.However,there are some uncertainties in the segmentation edge with deep semantic segmentation networks for rice mapping based on mid-resolution Landsat data,which limits the improvement of mapping accuracy.At the same time,the lack of high-quality large-scale training samples limited the application of deep learning technology to a certain extent.This paper takes Heilongjiang Province and the North China Plain(NCP)as the research areas,and conducts three aspects of research work based on Landsat data:(1)proposing a method for large-scale deep learning sample generation to construct highquality rice mapping training samples,and the presented method makes up for the lack of large-scale deep learning training samples;(2)constructing a rice mapping deep semantic segmentation network with high edge segmentation accuracy for Landsat data,and this network overcomes the defect of low edge segmentation accuracy and improves the mapping accuracy of the semantic segmentation network;(3)exploring the impact of Landsat images in different phenological periods on the accuracy of rice mapping network,and to provide a reference for the training samples selection of large-scale high-precision mapping.The main contents and conclusions are as follows.(1)In Heilongjiang Province and the NCP,the methods for making rice mapping training samples of deep semantic segmentation network based on Landsat data were respectively proposed,obtaining the high-quality large-area rice mapping training samples.Specifically,a total of 72 and 54 Landsat 8/9 OLI were respectively used to build large-scale training samples in Heilongjiang Province and the NCP.The workflow combining machine learning classification and manual correction was established for sample production in Heilongjiang Province.In the NCP,a patch-based region growing rate algorithm was proposed to combine mature period LSWI(Land Surface Water Index)and Sentinel 1 data,to improve the accuracy of the temporary mask of rice.Compared with the classical rice mapping based on transplanting and flooding period detection,this method in the study significantly reduced omissions.The accuracy validation showed that the F1 score of the training dataset was 0.980,and the overall accuracy was 0.978 in Heilongjiang Province,and the F1 score and overall accuracy of the training dataset in NCP were 0.949 and 0.929,respectively.This indicates that the dataset has high accuracy after manual correction,and can be used for training deep learning model.(2)The study proposed the Multi-Resolution Feature Fusion Unit(MRFU)and built the FR-Net deep learning semantic segmentation network based on MRFU.The MRFU structure consists of residual units and integrates multi-resolution feature flows,to maintain high spatial precision information of the network and realize the perception of multi-scale features.FR-Net is composed of multiple MRFUs,realizing the transmission and retention of high-resolution information flows in the network.The verification analysis showed that FR-Net had the best accuracy among the eight test models,with MCC and F1 scores of 0.893 and 0.898,respectively.FR-Net is less sensitivity to feature bands of Landsat for the rice mapping,the MCC only decreased by 4.14% as the feature band is not used,while the MCCs of other models,such as U-Net and RS-Net,decreased 16.84% and 11.11% respectively.(3)The difference in phenology between the training and testing images is an important factor affecting the mapping accuracy of deep semantic segmentation networks.When the imaging interval between the training and testing images was greater than 16 days,the relative change rate of the F1 score of the testing images in the study area was greater than 10%.Compared with spatial texture semantic information,spectral information is the main feature quantity for rice mapping using FR-Net and Landsat data.The impact of phenological differences on mapping accuracy is mainly due to changes in spectral information differences,with a relatively small impact on spatial texture semantic information.In Heilongjiang Province,the mean F1 scores of the images with and without spatial texture semantic information retained were 0.79 and 0.80,respectively.In the NCP,the mean F1 scores of the images with and without spatial texture semantic information retained were 0.722 and 0.701,respectively,and the F1 score of the accuracy of the images with spatial texture semantic information retained was improved by2.95%.To achieve higher accuracy of rice identification in NCP,training samples at a denser time scale are needed.(4)Using the cloud computing platform,large-scale rice monitoring and mapping were carried out based on the trained FR-Net and Landsat data.The rice planting area monitored in Heilongjiang Province were 51.55,51.95,55.52,54.84,66.45,60.56,57.35,and 63.93 million mu,respectively from 2013 to2020.In Heilongjiang Province,the rice planting area showed an increasing trend from 2013 to 2020,reaching a maximum area in 2017 and gradually declining thereafter.The rice planting area monitored in the NCP from 2013 to 2020 changed relatively stable,with rice planting areas of 22.07,17.51,17.85,18.51,21.04,20.84,21.07,and 20.37 million mu,respectively.