基于Sentinel-2A数据的森林覆盖变化研究

Research on forest cover change based on Sentinel-2A data

【目的】利用Sentinel-2A卫星重访周期短和波段信息丰富的特点,精确高效地获取森林覆盖变化信息,提高森林覆盖变化监测的时效性和精度。【方法】采用Sentienl-2A遥感影像作为数据源,以沅江市为研究对象,结合实地调查数据,选取地物训练样本,利用各地物的光谱指数特征和纹理特征来构建决策树模型进行地物分类,光谱指数包括NDVI、NDWI、NDBI和光谱反射曲线,纹理特征包括均值、方差、信息熵和对比度。对沅江市2016年8月1日和2017年5月18日的两期Sentinel-2A遥感影像数据进行地物分类,计算各地物面积并将两期分类结果中的森林覆盖区域提取出来,分析森林覆盖变化情况。【结果】1)利用光谱指数特征结合纹理特征构建决策树模型对沅江市进行地物分类,其地物总体分类精度为83.62%,Kappa系数为0.8257,森林制图精度为84.28%,森林用户精度为82.43%,比最大似然法的总体分类精度提升了11.27%,Kappa系数提高了0.133,森林制图精度提高了10.59%,森林用户精度提高了9.58%。2)在2016年8月1日至2017年5月18日期间沅江市森林面积减少了771 hm^2,有854 hm^2森林变为耕地,589 hm^2森林变为芦苇地,412 hm^2森林变为建设用地,105 hm^2森林变为水体。另外,有636 hm^2耕地、257 hm^2芦苇地、243 hm^2建设用地和53 hm^2水体变为森林。【结论】基于Sentienl-2A遥感影像数据,利用光谱指数特征和纹理特征构建决策树模型进行分类,能够有效提升地物分类精度;同时能够提高森林覆盖变化监测的时效性和精度,较为准确地分析森林覆盖变化情况,可为洞庭湖流域地物分类和森林覆盖变化监测提供决策支持。

【Objective】Based on the characteristics of short revisit period and rich band information of Sentinel-2A satellite,forest cover change information can be obtained accurately and efficiently,and the timeliness and accuracy of forest cover change monitoring can be improved.【Method】Sentienl-2A remote sensing image is used as the data source,and Yuanjiang city is taken as the research object,combines with actual survey data and selects training samples of ground object which uses the spectral index and texture features of each ground object to build a model of decision tree.The spectral index includes NDVI,NDWI,NDBI,and spectral reflection curve.In addition,the texture features include mean,variance,entropy and contrast.This paper classifies the two phases of Sentinel-2A remote sensing image data on August 1,2016 and May 18,2017 in Yuanjiang city,calculates the area of ground object,extracts the forest coverage area from the two phases of classification results,and analyzes the change of forest coverage.【Result】1)Using spectral index features and texture features to build decision tree model to classify the ground objects in Yuanjiang City.The overall classification accuracy is 83.62%,the kappa coefficient is 0.8257,and the forest mapping accuracy is 84.28%.The accuracy of forest users is 82.43%,which is 11.27%higher than the overall classification accuracy of the maximum likelihood classification,the Kappa coefficient is increased by 0.133,the accuracy of forest cartography is increased by 10.59%,and the accuracy of forest users is increased by 9.58%.2)Between August 1,2016 and May 18,2017,the forest area of Yuanjiang city decreased by 771 hectares,there are 854 hectares of forest in Yuanjiang becomes arable land,589 hectares of forest becomes reed land,412 hectares of forest becomes construction land,and 105 hectares of forest becomes water bodies.In addition,636 hectares of arable land,257 hectares of reed land,243 hectares of construction land and 53 hectares of water have been converted into forest.【Conclusion】Based on Sentienl-2A remote sensing image data,makes use of the spectral index features and texture features to build the model of decision tree model for classification can effectively improve the accuracy of feature classification;Meanwhile,it can also improve the timeliness and accuracy of forest cover change monitoring,more accurate analysis of forest cover change,and provide decision support for the Dongting Lake basin ground object classification and forest cover change monitoring.