遙測影像土地利用分類不確定性之探討

Abstract

遙測影像土地利用(Land-use/land-cover, LULC)分類被廣泛應用於地球表面監測、環境變遷偵測與水資源管理等領域。分類後常以分類主題圖呈現遙測影像分類的結果，並提供分類主題圖之分類正確率。遙測影像分類中監督式分類法需以訓練資料建立分類規則。由於訓練資料為從遙測影像選取的樣本資料，訓練資料的選取會造成分類正確率的不確定性。傳統上多數研究評估分類正確率時，會從遙測影像中獨立選取一組參考資料，對其分類建立混淆矩陣(confusion matrix)計算分類正確率。然而採用不同的參考資料可能求得顯著不同的分類正確率，且此分類正確率僅能代表以所選訓練資料建立之分類規則的分類效果，因此以一組參考資料混淆矩陣難以有效評估分類主題圖之分類正確率。此外，混淆矩陣呈現的分類正確率也會受到被分類資料的各類別抽樣數量影響。 以混淆矩陣評估分類效果的優勢在於可以呈現每個類別之分類情形，然而受到樣本選擇的影響，混淆矩陣呈現的分類正確率具有不確定性。本研究以多光譜遙測影像之序率模擬探討遙測影像土地利用之分類正確率評估，包括類別間樣本數比例對混淆矩陣的影響、訓練資料與參考資料選取造成的分類不確定性，與量化分類不確定性的方法。研究結果顯示(一)類別間樣本數比例對於分類正確率的評估具有決定性的影響 (二)訓練資料與參考資料的選擇會影響分類正確率的評估，而拔靴法訓練資料分類正確率可以做為全幅影像分類正確率(global accuracy)的估計(三)以一組訓練資料求得之拔靴法(Bootstrap)信賴區間，可做為全幅影像分類正確率的95%信賴區間。 對於遙測影像土地利用分類之不確定性，本研究總結以混淆矩陣評估分類正確率時，各類別樣本佔所有資料比例應與遙測影像中各類別被觀察到的機率相同。監督式分類的分類正確率評估應考慮訓練資料選取造成的的不確定性，為量化分類主題圖中各類別的分類不確定性，可以一組訓練資料透過拔靴法重覆取樣，對各類別使用者、生產者分類正確率與總體分類正確率分別求得全幅影像分類正確率的95%信賴區間，以此方式描述全幅影像分類結果之正確率，可以有效控制估計錯誤的機會。

Land-use/land-cover (LULC) classification using remote sensing images have been widely applied for earth surface monitoring, environmental change detection, water resources management, etc. All supervised classification methods require using a set of training data to establish class-assignment rules for pixels of unknown classes. A confusion matrix (or error matrix) which summarizes classification results of the training data or an independent set of reference data is then used to assess the classification accuracies of individual classes. However, classification accuracies of the training or reference data presented in the confusion matrix are estimates of the true and unknown classification accuracies of the population, i.e. all pixels of individual LULC classes. The class-assignment rules are derived from training data which are samples of individual classes, thus classification accuracies are inherently associated with uncertainties, i.e. sample variability would cause the uncertainty of classification accuracy. Besides, the ratio between sample sizes of difference classes could affect the classification accuracy. Conventional way of evaluating the LULC classification results is to choose an independent set of reference data and use the decision rules established by the training dataset for reference data LULC classification, and then examine the reference-data-based confusion matrix. We argue that such practices have limited capability in providing meaningful assessments since different reference datasets can yield significantly different accuracies in the reference-data-based confusion matrix. In this study, we investigated the uncertainty in classification accuracies (including the producer’s, user’s and overall accuracies) by using a bootstrap resampling approach. Unlike most LULC classifications, the proposed approach can provide 95% confidence intervals for class-specific classification accuracies. We also showed the theoretical basis and demonstrated that the conventional reference-data-based confusion matrix has very limited capability in assessing LULC classification results. Therefore, we conclude that LULC classification results should be evaluated by assessing the uncertainty of training-data-based confusion matrix and confidence intervals of class-specific classification accuracies can be provided by using the proposed bootstrap resampling approach.