Spatio-Temporal Data Fusion Based On Convolutional Neural Network And Extraction Of Winter Wheat Planting Area

With the rapid development of remote sensing technology,the use of time-series images for crop classification and identification has been widely used,and the fast and accurate extraction of crop planting area is important for estimating crop yield and ensuring food security.In order to solve this problem,this study constructs a spatio-temporal fusion model based on convolutional neural network to predict the missing high spatial resolution images and generate high spatial resolution images in combination with real high resolution images based on the existing research.NDVI,DVI time series set,while extracting texture features,and using different machine learning classification methods to classify crops and extract winter wheat planting area,in order to provide a basis and reference for similar studies.The main research results are as follows.(1)To address the problems that remote sensing images are difficult to obtain both time and space and the accuracy of existing spatio-temporal fusion models is not high,this study constructs a spatio-temporal fusion model based on convolutional neural network.Firstly,in the feature extraction stage of high and low resolution images,the basic convolutional layer is used to construct;secondly,in the fusion stage of images,the multi-scale convolutional lateral widening network is firstly used to extract multiple scales of feature maps without increasing the depth of the network,and then the channel attention module and spatial attention module are added to enable the network to focus on the channel and spatial dimensions of images again,deepen the learning of important features in images,and improve the network’s performance in extracting features.The network is then able to focus on the channel and spatial dimensions of the image again,deepen the learning of important features in the image,and improve the reconstruction ability of the network for extracted features.The model was trained and validated using the dataset constructed in this study,and better results were achieved in all metrics,with SSIM of 0.849,improving 0.03 and 0.028,respectively,compared with the results of STARFM and DCSTFN models,PSNR of 30.319,improving 0.827 and 0.768,respectively,in the prediction results for February 16,2020,ERGAS was 2.159,decreasing by 0.318 and 0.073,respectively,and RMSE was 0.034,decreasing by 0.002 and 0.001,respectively.From the visual effect,the texture and color are closer to the original image than the other two models.(2)Predictive reconstruction of missing images during the 2019-2020 winter wheat fertility period using a spatio-temporal fusion model,a 30 m 16-day medium-high spatial resolution NDVI and DVI timeset was constructed,and the timeset was smoothed and denoised using the S-G filtering algorithm,while the top 31 features were finally determined as the most optimal feature set based on the NDVI and DVI timesets,texture features,and other data,using the random forest feature preference method to rank the feature importance.Four of the top 10 important features in the most optimal collection are texture features,indicating that texture data play an important role in identifying crops.(3)The study established five identification schemes,NDVI timing,NDVI timing + DVI timing,NDVI timing + texture,NDVI timing + DVI timing + texture and feature preference,which were combined with support vector machine and random forest classification methods to classify winter wheat and other crops in the study area.The overall classification accuracy relationship is: Feature Preference > NDVI+DVI+Texture > NDVI+ DVI > NDVI+Texture >NDVI Timing.The feature preference combined with random forest classification method showed the best performance with an overall classification accuracy of 95.34% and a Kappa coefficient of0.93,which improved the overall classification accuracy by 3.85%,2.36%,1.94%,and 1.06%,respectively,compared with other schemes.The final extracted winter wheat area of 444509.10 mu has a relative error of about 2.4% with the actual area in 2020,and the relative error is reduced by 2%-4% compared with the fused single view image and Landsat image,and the results are reliable.Therefore,the images predicted by the spatio-temporal fusion model are used to construct multiple index time series sets,which can be used for effective extraction of winter wheat planting area in the county by combining with texture features.