Difference Between GF-6 WFV And Sentinel-2B In Land Use Classification And Vegetation Information Extraction

Land use and vegetation information is an important content and hot field of global environmental change and sustainable development.With the improvement of the resolution of satellite sensor,the feature information of the image is more and more abundant.The land use classification is more difficult because of the different types of ground objects and the obvious details and texture features.Therefore,how to extract these features optimally and apply different classification methods to obtain accurate ground feature information is particularly important.This study takes Hunyuan County of Shanxi Province as an example,taking the GF obtained on August 21,2019 as an example WFV and Sentinel-2Bimages acquired on September 3,2019 were used as the comparative data sources.The land use classification was carried out based on the rotating forest(Ro F)classification with feature space optimization(FSO).The classification results were compared with the random forest(RF),support vector machine(SVM),neural network(NN),decision tree(CART)classification and ROF classification without feature space optimization under the same conditions analysis.At the same time,vegetation coverage was calculated based on NDVI,Savi,redndvi and ndvi-dfi to further explore the vegetation extraction ability of different spatial resolution images in different research targets.The conclusions are as follows:(1)GF-6 WFV uses feature space optimization(FSO)algorithm to screen 35 optimal classification features,Sentinel-2B to screen 31 optimal classification features The overall classification accuracy of WFV and Sentinel-2B based on feature space optimization is 94.11%and 95.17% respectively,and the kappa coefficient is 0.9205 and 0.9301 respectively,which are higher than the ROF classification accuracy(92.06% and 93.81%)and kappa coefficient(0.8903 and 0.9047)before feature space optimization.Therefore,the feature space optimization algorithm not only reduces the dimension of feature variables,but also improves the image classification accuracy-The overall classification accuracy of 6 WFV and Sentinel-2B using ROF classification method was 94.11% and 95.17% respectively,and the kappa coefficients were 0.9205 and 0.9301 respectively,which were higher than the corresponding RF(91.24%,94.03%,0.8858 and 0.9209),NN(90.02%,89.76%,0.8762 and 0.8751),SVM(88.57%,90.04%,0.8639 and 0.8724),cart(87.29% and 88.75%),The results show that the classification accuracy of the classifier is better than that of the other classifiers.(2)On the whole,except that the overall classification accuracy and kappa coefficient of Gf-6 WFV based on NN are higher than those of Sentinel-2B,the overall classification accuracy(95.17%,94.03%,90.04%,88.75%)and kappa coefficient(0.9301,0.9209,0.8724,0.8649)of Sentinel-2B based on the other four classifiers(ROF,RF,SVM,cart)are higher than those of GF-6 WFV classification accuracy(95.17%,94.03%,90.04%,88.75%;0.9205,0.8858,0.8639,0.8564);from the classification accuracy of each feature,GF-6 Compared with Sentinel-2B,WFV images with lower resolution are easier to form mixed pixels,which makes Sentinel-2B’s classification accuracy of other five land features(cultivated land,forest land,grassland and shrub,bare land,urban and rural land)higher than GF-6 WFV images.However,according to other data,GF-6 WFV is closer to the real value for the large area of cultivated land and forest land.Therefore,Sentinel-2B image has better classification effect for grassland and shrub,bare land,urban and rural land which are broken,with obvious details and texture features,while GF-6 WFV is more suitable for large area distribution.(3)There were differences between GF-6 WFV and Sentinel-2B vegetation index.Among them,the NDVI,Savi and redndvi of GF-6 WFV in urban area were 4.30%,5.60% and 3.40%lower than those of Sentinel-2B,respectively,while those in forest area were 5.00%,2.30% and1.20% higher than those in Sentinel-2B.Sentinel-2B has higher NDVI standard deviation and more vegetation information,which is because it is easier to obtain small vegetation information from high-resolution images.The vegetation coverage ratios of Sentinel-2B in urban target area were 65.9549%,65.7168% and 62.4993% respectively,and the corresponding GF-6 WFV was difficult to extract sparse and scattered vegetation in buildings(61.6634%,61.8882% and60.6810%);the vegetation coverage ratios of Sentinel-2B in forest area were 90.3899%,90.3164% and 71.7005%,respectively WFV is difficult to distinguish the bare soil and road in forest(96.2454%,96.2725%,77.8173%).(4)The vegetation coverage of Sentinel-2B forest area extracted by ndvi-dfi model(89.1104%)and that extracted by pixel dichotomy method(ndvi-fvc-90.3899% and savi-fvc-90.3164%)were close,and the vegetation extraction results of urban area(60.5410%)were lower than that of pixel dichotomy model(ndvi-fvc-65.9549% and savi-fvc-65.7168%),but the difference was not significant The pixel dichotomy model can distinguish vegetation and bare land better,which is more in line with the actual situation.(5)The vegetation information extracted from different resolution images will show different differences in different ecosystems.Based on the vegetation coverage of low,medium and high vegetation coverage,the differences of NDVI,NDVI,NDVI and NDVI were relatively large,which were 83.06%,-2.06% and-2.25%,respectively.In the forest area,the difference of vegetation based on NDVI and Savi is the largest in the high coverage level,which is 11.38% and 11.61% respectively.The difference of vegetation coverage based on renndvi is larger in the low,medium and coverage level,which is-3.08%,-2.8% and 5.32% respectively.