應用監督式類神經網路於遙測影像分類

Abstract

監督式類神經網路如多層感知器、輻射半徑基底函數網路與支撐向量機，常用於遙測影像分類。但傳統上利用倒傳遞學習法訓練網路，可能會有推廣能力不足以及陷入局部、而非全域最佳解的問題。越複雜的問題需要越大型的網路來求解，上述問題也同時變得更明顯，因此尋找足以求解的最小規模網路是類神經網路的重要議題。 修剪網路先訓練一個足以求解但大於必要的傳統類神經網路，然後刪除其中不必要的連結，是獲得最佳網路架構、增進推廣能力的方法。但修剪網路的訓練階段一樣使用倒傳遞學習法，無法解決陷入局部最佳解的問題。 我們使用遺傳演算法，它是一種全域最佳化演算法，來實做演化式類神經網路。一般演化式類神經網路忽略交配運算，所以我們提出一個新的編碼方式來敘述網路架構與連結權重值，以網路的一個連結為單位，將與相同隱藏層節點連結的參數排在一起，使得交配運算得到的子代個體能有效繼承親代個體代表的網路架構特色。將一個大型網路所有連結的權重值以及存在與否等參數編碼，經由迭代演化求解，得到的最佳個體即為我們所求的最佳架構小規模網路。 實驗部分，我們首先利用XOR互斥問題表現使用新編碼方式的演化式類神經網路獲得最佳網路架構的能力；然後應用在遙測影像的分類上，與傳統多層感知器、輻射半徑基底函數網路、kernel SVM、修剪網路做比較。結果顯示：演化式類神經網路同時具備優秀的重現能力與推廣能力，並且能最快收斂、獲得架構最小的網路。

Supervised neural networks (NN) as multilayer perceptron (MLP), radial basis function network (RBFN), and support vector machine (SVM), are usually used for remote sensing image classification. But there may be some drawbacks of the back-propagation (BP) learning algorithm for training. These drawbacks include badly generalizing and converging not to its global optimal solution but to a local one. When we want to approximate a complex problem, we will need to use a large NN with a lot of hidden nodes, and make the above mentioned drawbacks more evident. In NNs, as in all modeling problems, to search the simplest network that can solve problem enough is a serious issue. Pruning network eliminates certain unnecessary links after a lager immoderately traditional BPNN has been trained. This is a method to obtain optimal architecture of NN with generalization raised. But another drawback, converging to local optimal solution, still exists because pruning network is trained with BP learning algorithm. Genetic algorithm (GA) is one of global optimization algorithms. We use GA to achieve evolving artificial NN (EANN). Unlike EANNs ignoring crossover operation in general, we propose a new encoding method to describe all links exist or not and their weight coefficients in a NN. This method arranges parameters about a same hidden node together, so child individuals produced by crossover operation can inherit characters about network architectures from their parents effectively. Encode information of a large NN, evolve through some generations, and then obtain optimal architecture of NN by decoding the optimal individual. In the experiment, we use XOR problem to show performance of EANN with new coding method. Then we apply it to do remote sensing image classification, comparing with traditional MLP, pruning network, RBFN, and kernel SVM. These results show that EANN has both high repeating ability and good generality, converges fastest, and obtain an optimal architecture of NN with smallest amount of network links.