大气校正对SPOT卫星遥测水质的影响

Effect of Atmospheric Correction on Stream Water Quality Monitoring by Using Spot Satellite Remote Sensing Images

藉由卫星遥测进行河川水质监测,目前尚没有较明确可行之方法,如何利用较为简单且适当的SPOT卫星遥测大气校正方法,正确辨识水体水质,是本研究的主要目的。利用SPOT卫星作两阶段非监督式及监督式自动分类,确认卫星影像中水质测站对应之水体样本,并将所有样本依季节分群,俾让卫星监测水体水质样本较为均质。模拟方式采用多变量回归、类神经网络及判别分析3种模式,并比较4种不同之大气校正程序。结果发现,以水质及其指标整体预测来看,类神经网络预测结果较优于多变量回归及判别分析的结果,大气校正方法以直接采用灰度值并消除最暗像元灰度值之校正方法,即可达到不错之预测结果。综合而言,以SPOT或分辨率更高之卫星光谱遥测水质是简单可行,但仍需更多数据以验证其精确度。

By using the remote sensing data we can carry on the stream water quality detection. Most successful studies on water quality monitoring by remote sensing were mainly relied on choice of feasible method of atmospheric correction. This research incorporates digital count, radiance, reflectance, and reflectance with transmittance four different correction procedures to evaluate the effect on simulation. Dark object subtraction was selected for all procedures, and followed by separating the samples into two groups for the reason of the seasonal variation. In order to consider the sampling difficulty on SPOT images with its limited pixel resolution, two step unsupervised pre-classification and supervised post-verification were used for extracting the reliable water pixels from SPOT images which are corresponding to the same water quality monitoring locations in ten different days. The study adopted multivariate regression （MR）, artificial neural network （ANN）, and discriminant analysis （DA） to examine and compare the results of different relationships between optical spectrum and water quality. The overall results showed that the analysis from multivariate regression and discriminant analysis were not as good as the results obtained from artificial neural network in the study area. For atmospheric correction, the simple digital count with dark object subtraction method is necessary and sufficiently enough to count on atmospheric interference by comparing the results from four different correction procedures. However, this result of limited optical data correction and learning technique needs to be further confirmed by using higher resolution satellite images and more case studies. Basically, it is evident that artificial neural network has the potential and feasibility of monitoring water qualities and its derived index.