Study On Fast Extraction Of Forest And Non-forest Based On GaoFen-6 The Wide Field Of View Camera Data

With the increase of domestic high spatial resolution satellite,more and more remote sensing data have been used in forest surveys,forest management,and forest change detection.GaoFen-1(GF-1)satellite has shown some potentials in forest information retrieval,but its limitation in spectral resolution,containing only 4 spectral bands,restrict the capacity for mapping forest and non-forest.In June 2018,China launched Gaofen 6(GF-6)satellite,the first self-owned widefield-of-view,multi-spectral satellite with red and yellow edge bands,which could contribute a lot on forest and non-forest mapping.With the support of the common key technology project of national highresolution earth observation system,this study first evaluates the reliability of the three global forest product through area consistency and spatial consistency indicator in typical regions of China,including Tianshui City,Huangshan City,Yichang City,Pu'er City,Chengde City,Benxi City,Fangchenggang city,based on national authorized forest distribution map.Then,selecting Huangshan City and Tianshui City as pilot sites,the potential of GF-6 with different machine learning algorithms on forest and non-forest rapid identification,including K nearest neighbor(KNN),naive Bayes(NB),CART decision tree(CART),and support vector machine(SVM),Random Forest(RF),were explored through the comparisons of different strategies,with a special focus on GF-6's new band,and multi-temporal features.Finally,the model-based and statistical methods are used to evaluate the importance of input factors,and the causes of the differences in classification accuracy are analyzed through the analysis of spectral characteristics of the detailed forest land type.The main conclusions are as follows:(1)The reliability assessment results indicate that the reliability of global forest products in the southern region is better than that in the north;the reliability of FROM-GLC 10 data products is better than that of FNF and GFCM data product.(2)Through the comparison of five machine learning classification methods,the random forest classification method has the highest overall classification accuracy in the forest and non-forest land classification in Huangshan and Tianshui research areas,followed by support vector machines,and the other three classifiers exist big differences.(3)GF-6 red edge band,red edge vegetation index,multi-temporal feasures and terrain features have important potential for the classification of forest and non-forest,especially for non-typical shrubs.(4)In Huangshan and Tianshui,sparse forests including tea trees,apple trees,walnut trees,pepper trees,and mulberry trees are very similar to the spectral characteristics of cultivated land types,which is one of the main causes of forest and non-forest classification errors.