基于神经网络拟合多源遥感信息估算冠层高度

Estimation of forest canopy height using neural network fitting multi⁃source remote sensing information

为了获取空间上连续的森林冠层高度分布,借助后向传播(BP)神经网络,拟合2019-2020年冰、云和陆地高度卫星(ICESat-1)的陆地/植被产品ATL08的生长季树高数据和中分辨率成像光谱仪(MODIS)的地表信息等多源遥感资料,探索我国西南地区森林冠层高度的空间连续估算。结果如下:(1)ICESat-2/ATL08遥感数据对西南地区森林冠层高度表达可信,与2019年同期其他树高遥感数据对比,通过了置信度检验(R=0.566,P<0.01)。(2)BP神经网络拟合ICESat-2/ATL08和MODIS数据得到西南地区空间连续森林冠层高度的方法可靠。神经网络训练结果决定系数(R2)达0.7以上,均方根误差在2.5 m左右。从结果中随机选择10个自然保护区位置森林冠层高度数值,对比2005年同源遥感资料,得到R=0.836(P<0.001)。连续森林冠层高度数值的频率直方图符合正态分布,高频段为10~25 m。(3)估算得到的西南地区森林冠层高度平均值为16 m,低于2005年遥感均值,但在重庆与贵州交界、贵州东北部、云南六诏山等地,森林覆盖度和冠层高度均有所提升,云南大理的则略有减少。从空间分布来看,大于20 m的数值分布于四川盆地外围、金沙江流域、重庆大巴山、云南哀牢山和无量山以及贵州东南部等地区;四川西南部乡城和稻城、重庆东北部、云贵高原中部地区均在10 m以下。

Estimating continuous spatial forest canopy height is difficult in forestry investigations.Using a back propagation(BP)neural network,the ice,cloud,and land elevation satellite(ICESat⁃2)and its terrain/vegetation products[ICESat⁃2/ATL08 data in growing seasons and moderate resolution imaging spectroradiometer(MODIS)information in 2019-2020]were combined to explore the continuous spatial estimation of forest canopy height in southwest China.The results are as follows:(1)ICESat⁃2/ATL08 data reliably expressed the natural forest canopy in southwest China and passed the confidence test with R=0.566,P<0.01,compared with remote sensing data for tree height during the same period in 2019.(2)BP neural network fitting of ICESat⁃2/ATL08 and MODIS data could reliably obtain continuous forest canopy height.Neural network training results showed R2>0.7 and root mean square error of about 2.5 m.The forest canopy height values of 10 randomly selected natural reserves were compared with the homologous remote sensing data from 2005,resulting in R=0.836(P<0.001).The frequency histogram of continuous forest canopy height showed normal distribution,and the high frequency band was 10-25 m.(3)The mean forest canopy height in southwest China was 16 m,lower than the mean in 2005.However,the forest coverage and forest canopy height increased on the border between southern Chongqing and northern Guizhou,northeastern Guizhou,and Liuzhao Mountain in Yunnan,and it decreased slightly in Dali City in Yunnan.In terms of spatial distribution,values greater than 20 m were distributed in the periphery of Sichuan Basin,the Jinsha River Basin,and Daba Mountain in Chongqing,Wuliang Mountain and Ailao Mountain in Yunnan,and southeastern Guizhou.Rural towns in southwest Sichuan,Daocheng,northeast Chongqing,and the central Yunnan⁃Guizhou Plateau were all below 10 m.