基于GF-1 WFV影像和BP神经网络的太湖叶绿素a反演

叶绿素a浓度是可直接遥感反演的重要水质参数之一,常用来评价湖泊水体的富营养化程度.太湖是典型的二类水体,光学性质复杂,应用一类水体线性反演模式拟合较为片面且难以找到最佳拟合模型.BP神经网络模型具有模拟复杂非线性问题的功能.为研究高分一号卫星16m多光谱相机WFV4结合BP神经网络进行太湖叶绿素a浓度监测的可行性,实验利用GF-1 WFV4影像和实时的地面采样数据,建立了BP神经网络模型,同时采用波段比值经验模型进行对比.经精度检验,BP神经网络模型预测值与实测值之间的可决系数R2高达0.9680,而波段比值模型的R2为0.9541,且均方根误差RMSE由波段比值模型的18.7915降低为BP神经网络模型的7.6068,平均相对误差e也由波段比值模型的19.16%降低为BP神经网络模型的6.75%.结果证明,GF-1 WFV4影像应用BP神经网络模型反演太湖叶绿素a浓度较波段比值模型精度有所提高.将经过水体掩膜的GF-1 WFV4影像用于训练好的BP神经网络反演太湖叶绿素a浓度分布,结果显示,叶绿素a高浓度区集中分布在湖心区北部、竺山湾、梅梁湾区域,与之前的研究一致.本文研究结果验证了采用BP神经网络模型对GF-1 WFV4影像进行太湖叶绿素a浓度反演的可行性.

The study of inversion of chlorophyll a in Taihu based on GF-1 WFV image and BP neural network

Chlorophyll a, which is commonly used to evaluate the eutrophication of limnetic water, is one of the most important water quality parameters that can be directly retrieved by remote sensing. As Lake Taihu is the typical second class water body with complex optical characteristics, it is difficult to find the applicable linear inversion models usually applied to the first class water body. To evaluate the applicability of retrieving chlorophyll a from multispectral camera GF-1/WFV4 with 16 m spatial resolution, the BP neural network model was established due to its advantage that can simulate complex nonlinear problems, and applied to Lake Taihu combined with quasi-synchronous ground sampling data. The retrieval result was subsequently compared with that of the empirical model of band ratio. The R² between BP model predicted values and measured chlorophyll a is as high as 0.9680, while the band ratio model R² is 0.9541. Precision comparison between BP neural model and band ratio model shows that the root mean square error RMSE is reduced from 18.7915 of the latter to 7.6068 of the former, besides that the average relative error is reduced from 19.16% to 6.75%. It is proved that the BP neural network model is better to obtain the chlorophyll a than the band ratio model in Lake Taihu using GF-1 WFV4. After land-water masking, the GF-1 WFV4 image was taken into the trained BP neural network to retrieve the spatial concentration distribution of chlorophyll a over Lake Taihu. The result shows that high chlorophyll a concentration distributes in the north lake, Zhushan Bay and Meiliang Bay area, which is consistent with the previous studies. The result validates the feasibility of using GF-1 WFV4 image to retrieve the chlorophyll a concentration in Lake Taihu with BP neural network model.