Estimation Method And Spatio-temporal Continuous Mapping Of Fractional Vegetation Cover With Landsat Data

Fractional vegetation cover(FVC) is an important parameter for studying the dynamic process of vegetation ecosystem on earth's surface,and has an important effect on the ecological environment,climate modeling,and carbon dioxide circulation.In addition to small-scale ground measurement,the estimation of FVC is currently based on satellite data such as MODIS(Moderate Resolution Imaging Spectroradiometer)and Landsat.Low spatial resolution vegetation coverage products such as VGT bio GEOphysical product Version2(GEOV2),PROBA-V bio GEOphysical product Version 3(GEOV3),and Global LAnd Surface Satellite(GLASS)range from a few kilometers to hundreds of meters.Although they have good temporal and spatial continuity,they do not meet the fine-scale regional analysis.Therefore,we outline an effective framework to produce Landsat-like spatial resolution fractional vegetation cover products by combining multi-source satellite data(high spatial resolution(HSR)images(Gao Fen-2,GF-2)and Landsat data)and machine learning algorithm.First,we use the preprocessed GF-2 images with a spatial resolution of 1m and corresponding Landsat images to extract sun angle and Landsat band reflectance information as a reference for Landsat-8 to calculate vegetation cover.After a large number of training samples are obtained,they are used to train a random forest(RF)regression model.The model accuracy is evaluated by random selection of training samples,and the root mean square error(RMSE)is less than 0.1,and R~2>0.9.At the same time,the FVC obtained by this research is compared with the Imagine S site around the world,the result shows RMSE<0.2 and R~2>0.8,which indicates that the model in the paper has a high accuracy.The inter-comparison with the GLASS FVC product at a scale of near 500m(R~2>=0.9)confirm the effectiveness of our model and also indicate the high consistency of two data.In addition,compared with the FVC extracted by the supervised classification algorithm from the ultra-high-resolution unmanned-aerial-vehicle(UAV)images,the R~2>0.7.These high precision results show that the extraction method proposed in the study can be applied to large-scale and fine-grain fractional vegetation cover estimation.With the low time resolution of Landsat data and the problem of discontinuity in space due to influence of cloud or snow pollution,we attempt to reconstruct Landsat FVC data,arming to provide a more effective data support for a long time series FVC analysis and change detection.