基于机器学习方法的太湖叶绿素a定量遥感研究

为了比较评价人工神经网络和支持向量机2种机器学习算法在水质遥感中的应用能力，本研究首先从基础理论和学习目的入手，对比分析了2种机器学习算法的理论体系；其次，以太湖为例，基于MODIS遥感影像，构建了反演太湖叶绿素a浓度的2种机器学习方法模型，通过对模型的验证、稳定性和鲁棒性分析以及全湖反演结果对比3个方面评价了2种模型的泛化能力.验证结果表明，支持向量机模型对验证样本预测结果的均方差根和平均相对误差分别为5.85和26.5%，而人工神经网络模型的预测结果均方差和平均相对误差则高达13.04和46.8%；稳定性和鲁棒性评价亦说明，以统计学习理论为基础的支持向量机模型具有更加良好的稳定性、鲁棒性，空间泛化能力优于人工神经网络模型；2种机器学习算法对太湖叶绿素a的浓度分布反演结果基本一致，但人工神经网络模型因其学习目标设定和网络构建中的“过学习”等缺陷，造成了对东太湖以及湖心区叶绿素a的反演结果与实际监测结果差异较大.

Quantitative Retrieval of Chlorophyll a Concentration in Taihu Lake Using Machine Learning Methods

We evaluated the performance of two machine learning methods, artificial neural net (ANN) and support vector machine (SVM), for estimation of chlorophyll a in Taihu Lake from remote sensing data. The theoretical analysis has been done from basic theory and learning target of these two methods first. Then two empirical algorithms have been developed to relate reflectance of MODIS to in situ concentrations of chlorophyll a. The performance of ANN and SVM is comparatively analyzed in terms of validation, stability and robustness assessment and chlorophyll a distribution of Taihu Lake from two algorithms. The root of mean square deviation (RMSE) and average relative error (ARE) of validation data is only 5.85 and 26.5% of SVM retrieval model, however, RMSE and ARE of ANN model is 13.04 and 46.8%. Stability and robustness assessment suggest that SVM provides the better performance than ANN. And the retrieval results show that the chlorophyll a distribution of the whole lake from two algorithms is similar, however, the chlorophyll a concentration in the eastern region and central region of Taihu Lake is distorted by ANN model because of the limitations, such as learning target setting and over-learning in net construction.