在農業空拍影像辨識中，高品質的資料集（例如精確標註的空拍地塊資料）對 於達成可靠的模型效能與評估至關重要。先前的研究 [1] 主要聚焦於空拍影像中的異 常值偵測，但也發現其在測試階段能有效辨識錯誤標註的地塊。然而，當面對刻意錯 誤標註的樣本時，該方法無法偵測到其中相當大部分的案例，顯示其在錯誤標註資料 偵測上的魯棒性存在限制。此外，農業影像分類常涉及由多個外觀相似的子類別所組 成的類別，這種特性會對深度學習模型正確分類單一地塊造成負面影響。 為了應對這些挑戰，本研究提出一種新的基於特徵的錯誤標註偵測方法，以提 升農業空拍影像的分類效能。受到異常值偵測技術的啟發，我們的方法結合了 Weibull 分佈建模與深度特徵擷取 [1]。我們首先從預訓練模型中擷取深度特徵，接著進行物理 表示分析，以揭示特徵空間中的結構性不一致。之後，我們應用一個資料淨化流程， 透過過濾疑似錯誤標註的樣本來精煉資料集。最後，藉由使用 Weibull 分佈來統計建 模特徵分佈的尾部，使得異常或錯誤標註資料的識別更加有效。 實驗結果顯示，所提出的方法不僅能與現有的錯誤標註偵測技術競爭，且在 召回率 與 精確率 上於空拍地塊資料集及細粒度的 CUB-200-2011 資料集皆取得更優異的表現。;In agricultural aerial image recognition, high-quality datasets such as accurately la- beled aerial imagery parcel data are important to achieve reliable model performance and evaluation. Previous research [1] mainly focused on outlier detection in aerial imagery, but it was also found to be effective in identifying mislabeled parcels during testing. How- ever, when tested with intentionally mislabeled samples, the method failed to detect a significant portion of these cases, highlighting limitations in its robustness for mislabel detection. Moreover, agricultural image classification often involves classes composed of visually similar subclasses, which can negatively impact deep learning models’ability to correctly classify individual parcels. To address these challenges, this study proposes a new feature based mislabel de- tection approach aimed at improving classification performance in agricultural aerial im- agery. Inspired by outlier detection techniques, our method leverages Weibull distribution modeling and deep feature extraction [1]. We begin by extracting deep features from a pre-trained model, followed by a physical representation analysis to uncover structural in- consistencies within the feature space. A purification process is then applied to refine the dataset by filtering out suspected mislabeled samples. Finally, the Weibull distribution is utilized to statistically model the tails of the feature distribution, enabling more effective identification of anomalous or mislabeled data. Experimental results demonstrate that the proposed method not only competes with existing mislabel detection techniques but also achieves superior performance in terms of recall and precision on both an aerial image parcel dataset and the fine-grained CUB-200-2011 dataset
