Research On Remote Sensing Image Classification Based On Optimized Spatio-temporal Fusion Model

High temporal resolution and high spatial resolution remote sensing data are of great significance for global or regional scale applications,such as land cover mapping,vegetation phenological change monitoring,agricultural crop yield estimation,and surface temperature monitoring.However,the current satellite remote sensing data is limited by hardware technology and development cost.The existing satellite remote sensing data cannot meet the requirements of high spatial resolution and high time resolution at the same time.That is,the time resolution and spatial resolution of satellite remote sensing data are large.In most cases,they cannot coexist.Aiming at the problem that it is difficult for a single satellite remote sensing data to coexist with high temporal resolution and high spatial resolution,this paper studies the problem of solving the contradiction between the spatial and temporal resolution of satellite remote sensing data by improving the spatio-temporal fusion model.The essence of the classic spatiotemporal fusion model is to apply the linear relationship between the two low spatial resolution images at the reference and prediction moments to the prediction of high spatial resolution images,and construct a weighting function to assist in obtaining spatiotemporal information.The core idea of the algorithm is to construct high spatial resolution images by extracting the time information of low spatial resolution images.In order to fully tap the application potential of high spatiotemporal resolution remote sensing data in actual production,this paper proposes an optimized spatiotemporal fusion model,which improves the accuracy of image prediction,thereby promoting the application of remote sensing spatiotemporal fusion technology in various fields.This article first studies the classic spatio-temporal fusion model,and proposes optimized spatio-temporal fusion model for its shortcomings.Based on the optimized spatio-temporal fusion model,the long-term remote sensing image is generated and the NDVI time series curve is extracted.The random forest classifier is used to classify the study area,and the classification results of a single image are compared and analyzed.The main research contents and conclusions of this paper are as follows:(1)Aiming at the problem of inaccurate selection of similar pixels in the rectangular neighborhood window of the classical spatiotemporal fusion model,this paper proposes Superpixel based Spatial and Temporal Adaptive Reflectance Fusion Model(S-STARFM).The classic spatio-temporal fusion model has poor stability in selecting similar pixels in a rectangular neighborhood window,and the accuracy of the selected similar pixels is not high.Pixels in the superpixel neighborhood window have higher spectral and texture similarity,so there are more similar pixels in the superpixel.It further proves the credibility and feasibility of superpixels as neighborhood windows.In this paper,according to the characteristics of similar features in the super-pixel neighborhood,the rectangular neighborhood window of the STARFM model is adjusted to the super-pixel neighborhood window to improve the accuracy of selecting similar pixels,thereby improving the accuracy of predicted images.Experiments with classic data sets found that the S-STARFM model achieved higher prediction accuracy,and the overall error index was significantly reduced.Experimental results prove that the SSTARFM model can predict more spatial information,while at the same time it can better maintain the image details and avoid color distortion.(2)Aiming at the problem that the classic spatiotemporal fusion model does not have high prediction accuracy in abnormally changing areas,this paper proposes Hierarchical based Spatial-Temporal Fusion Model(H-STFM).Taking into account the different changes of ground features,the spatio-temporal fusion framework is divided into two layers for prediction.First,the target pixels to be predicted are divided into phenological change pixels and mutation pixels,and different models are used for prediction respectively.The layering strategy can better distinguish and predict the changing pixels of different types of features,so that the predicted image is closer to the real image.The H-STFM model is qualitatively and quantitatively evaluated through experiments on classic data sets.In terms of visual effects,the predicted image of the H-STFM model is closer to the real image;the quantitative evaluation indicators show that each quantitative evaluation indicator of the H-STFM model is the optimal value,and the overall similarity between the predicted image and the real image The highest degree and the smallest error fluctuation.The layered fusion spatio-temporal fusion framework can better predict the changes of different ground object types,and further improve the prediction accuracy of the fusion image.(3)Based on the H-STFM model to predict the remote sensing image data in the study area,use Landsat and MODIS images as source data to generate long-term Landsat images.The NDVI time series curve is further extracted,and the phenological characteristics of various vegetations at different times are distinguished by analyzing the changes in the NDVI curve of different vegetations,and this is used as the basis for remote sensing image classification.Select samples through Google Earth map,and use random forest classifier for image classification.Comparing with the classification results of a single image,it is found that remote sensing images based on time series have higher classification accuracy and are more conducive to distinguishing different vegetation types.