基于客观阈值与随机森林Gini指标的水体遥感指数对比

Comparison of Remote Sensing Water Indices based on Objective Threshold Value and the Random Forest Gini Coefficient

利用福建福州、西藏尼玛和澳大利亚弗伦奇3地代表不同水体类型的Sentinel-2A MSI和Landsat-8 OLI数据,采用客观阈值法(0阈值)和随机森林重要性评估法,比较和分析了改进型归一化差值水体指数(Modified Normalized Difference Water Index,MNDWI)、自动水体提取指数(Automated Water Extraction Index,AWEI)和水体指数2015(Water Index 2015,WI2015)这3种世界常用的水体指数之间的差异。从水体增强的效果来看,MNDWI增强的水体不仅具有丰富的信息还具有鲜明的对比度,AWEI和WI2015增强的水体信息的对比度相对偏弱。精度验证表明:3种指数提取的水体精度都较高,但MNDWI在3个地区的平均总精度略高于WI2015和AWEI,3者的平均总精度分别为91.83%、91.16%和90.07%。在提取细小水体方面,MNDWI的能力强于其他2种指数,在阴影较为明显的高原山地区域,MNDWI提取水体的效果优于AWEI和WI2015。进一步采用随机森林的Gini指标进行的重要性评估表明,MNDWI在区分水体和非水体的分类中表现出了很强的重要性,尤其在Sentinel-2A MSI数据中表现得更为突出,而WI2015和AWEI的重要性则相对较弱。

This study used Sentinel-2A and Landsat-8 images of Fuzhou in Fujian,Nima in Tibet,China and French Island in Australia to assess the performance of three commonly-used water indices,i.e.,Modified Normalized Difference Water Index(MNDWI),Automated Water Extraction Index(AWEIsh and AWEInsh)and Water Index 2015(WI2015).The objective threshold value,i.e.,0 threshold,and random forest importance assessment method(Gini coefficient)were adopted to do the comparison with different water types(river,lake,and ocean).Among the water enhanced images of the three indices,MNDWI-enhanced water image has the highest contrast and rich information,whereas AWEI and WI2015 have relatively low contrast and are less informative.Accuracy validation shows that the water features extracted by the three indices all have high accuracy.Nevertheless,the average overall accuracy of MNDWI in the three areas is slightly higher than that of WI2015 and AWEI,which are 91.83%,91.16%and 90.07%,respectively.In addition,MNDWI can detect small water bodies and remove mountain slope shadows more effectively than the other two indices.The importance assessment revealed by the Gini coefficient of random forest further shows that MNDWI has the strongest importance in the separation of water with non-water features,especially shown in Sentinel-2A images,while WI2015 and AWEI have a relatively lower importance.