支持向量機及Plausible Neural Network於水稻田辨識之研究

Abstract

水稻是台灣地區重要作物，其種植面積調查是政府每年的重要工作之一，透過該面積調查可用於預估水稻產量，作為政府糧食政策擬定及災害補助之依據。自1980年起，台灣地區水稻面積之調查，主要是由糧食署依「台灣地區稻米生產調查改進要點」進行辦理，該作業是採航空照片並經人工判釋方式得水稻田種植面積，需耗費大量人力與物力。倘若能以衛星影像配合分類理論，建立自動化的辨識系統，將能有效減少成本的投入及人為辨識上的主觀差異。 本研究為以支持向量機及Plausible Neural Network進行水稻田辨識之研究，採用Formosat-2（嘉義實驗區）及SPOT（新竹實驗區）影像為資料來源，並和高斯最大似然分類法、倒傳遞類神經網路、學習向量量化類神經網路、輻狀基底函數類神經網路及粗糙集方法所得之分類成果進行比較。Plausible Neural Network並採用兩種版本進行分類實驗，V 1.3為一般化的版本；V 1.0則加入考量樣本屬性間及鄰近樣本間之資訊。支持向量機會因選用核函數的不同，而對分類成果造成差異，故兩實驗區皆進行支持向量機於不同核函數及各分類理論之辨識作業。研究中，並以嘉義實驗區資料，探討支持向量機與高斯最大似然分類法於不同群聚類別數及訓練樣本集合之穩定性及四種紋理影像生成方式輔助支持向量機分類所得之成果差異。 實驗所得成果以誤差矩陣作展示並進行精度評估。研究成果顯示，在兩個實驗區中，皆以Plausible Neural Network V1.0所獲得的分類成果為最佳，在嘉義實驗區之整體精度達94%以上；而在新竹實驗區之整體精度達91%以上。支持向量機之分類成果，在嘉義實驗區僅次於Plausible Neural Network V1.0；而在新竹實驗區則次於Plausible Neural Network V1.0及學習向量量化類神經網路。支持向量機於不同核函數分類實驗，以輻狀基底函數及多項式較線性及兩層式類神經網路適用於本研究，於嘉義實驗區以多項式之成果為最佳；而在新竹實驗區則以輻狀基底函數之成果為最佳。於不同群聚類別數及訓練樣本集合之分類實驗，以支持向量機的穩定性較高斯最大似然分類法為佳，其整體精度變動皆相對較高斯最大似然分類法為小，約在1%以內；而高斯最大似然分類法之整體精度變動，最大可達10%以上。紋理影像輔助支持向量機分類實驗，由成果顯示，紋理影像的加入對分類精度提升的成效有限，整體精度最高提升約0.3%，某些紋理影像的加入，甚至降低原有的分類精度，故紋理特徵的選取及使用必須加以考量。

Rice is the most important crop in Taiwan. Its field planting inventory is a routine government operation. The inventory is used to support decisions and policies related to crop yield estimation, hazard mitigation, and other relevant areas. Starting in 1980, the government has been utilizing aerial photo interpretation as the primary inventory method. This procedure is based on manual interpretation. If an automated classification system can be implemented, both the time and cost required for the inventory can be reduced. The errors caused by subjective interpretation of data by humans can also be avoided. This study investigates the application of the Support Vector Machine (SVM) and Plausible Neural Network (PNN) for image classification. The images used for the experiment include multi-temporal Formosat-2 images of the Chiayi area and multi-temporal SPOT images of the Hsinchu area. Other classification schemes are used for comparison: Gaussian Maximum Likelihood Classification, error Back-Propagation (BP) neural network, Learning Vector Quantization (LVQ) neural network, Radial Basis Function (RBF) neural network, and Rough Set theory. Two implementations of PNN are adopted. V1.3 is the general version, while V 1.0 takes both spatial and attribute relations into account. On the other hand, there are a number of kernel functions to be selected with SVM. The stability of SVM and Maximum Likelihood Classification, with respect to the contribution of texture images and the number of clusters and training samples, in the Chiayi dataset are also evaluated. PNN V1.0 performs the best for both datasets. The Overall Accuracy is higher than 94% for Chiayi and 91% for Hsinchu. SVM performs second-best for Chiayi and third-best for Hsinchu. Among different Kernel functions, RBF and polynomials are shown to be better than linear and two-layer neural networks. The Polynomial Kernel Function is the best for Chiayi and RBF is the best for Hsinchu. SVM has higher stability than Maximum Likelihood Classification with respect to the number of clusters and training samples. The contribution of texture images for classification is not significant. Instead, the influence of some texture images is found to be negative.