近年來，深度學習已廣泛應用於農地作物分布與生長狀態監測。然而，現有農業影像分析方法多仰賴監督式學習，需大量精確人工標記資料；對於高密度坵塊分布之航拍影像而言，逐一標記作物生長階段成本高昂。此外，作物生長具有連續漸進特性，相鄰階段間常僅存在細微紋理、色澤與局部結構差異，使無監督分群模型容易產生群 集混淆與分群不穩定。為降低人工標記依賴並提升細粒度生長階段分群能力，本研究提出一套適用於航拍農作物坵塊影像之無監督深度分群框架。所提方法以SeCu為基礎，採用弱資料增強以保留作物細部特徵，並提出餘弦相似度中心更新機制（CosineSimilarityCenter Update， CSC），透過高信心樣本篩選與代表性中心樣本更新，降低低信心樣本與邊界樣本對群集中心造成的影響。此外，本研究加入模組度最大化損失（Modularity Maximization Loss， MML）作為結構正則化機制，以強化群內聚合與群間分離。 實驗結果顯示，本方法於水稻資料集之ACC、ARI與NMI分別達0.925、0.795與 0.826；於六種作物多期程任務中，ACC達0.531，較次佳方法提升9.8個百分點；於十種作物大規模多期程實驗中，ACC達0.533，為所有比較方法中唯一超過0.5者。 此外，本方法於MH-SoyaHealthVision 葉片疾病資料集亦取得0.916、0.828與0.815之 ACC、ARI 與NMI，並在5-FoldCross-Validation 中維持穩定表現。整體而言，本研究方法能在無人工標籤參與訓練的條件下學習具分群鑑別力之作物生長特徵，並可作為 大量航拍坵塊影像於標記前輔助判讀與資料整理之前置參考。;Deep learning has been widely applied to cropland monitoring and crop growthassessment. However，mostagriculturalimageanalysismethodsrelyonsupervisedlearningandrequirelarge amounts of accurately annotated data. For high-density aerial field parcel imagery， manually labeling crop growth stages is costly and time-consuming. Moreover， crop growth is continuous and gradual， and adjacent stages often differ only in subtle texture， color， and local structural changes， making unsupervised clustering prone to cluster confusion and unstable assignments. To reduce the dependence on manual annotations and improve fine-grained crop growth stage clustering， this study proposes an unsupervised deep clustering framework for aerial crop parcel imagery. The proposed method is built upon SeCu and adopts a milder data augmentation strategy to preserve fine-grained crop features. A Cosine Similarity Center Update (CSC) mechanism is introduced to update cluster centers using representative high-confidencesamples， thereby reducing the influence of low-confidence and boundary samples. In addition， Modularity Maximization Loss (MML) is employed as a structural regularization term to enhance intra-cluster compactness and inter-cluster separability. Experimental results show that the proposed method achieves ACC， ARI， and NMI scores of 0.925， 0.795， and 0.826 on the rice dataset. In the six-crop multi-stage task， it achieves an ACC of 0.531， improving over the second-best method by 9.8 percentage points. In the ten crop large-scale multi-stage experiment， it achieves an ACC of 0.533 and is the only method with ACC exceeding 0.5. The proposed method also achieves ACC， ARI， and NMI scores of 0.916， 0.828， and 0.815 on the MH-SoyaHealthVision leaf disease dataset， and maintains stable performance under 5-Fold Cross-Validation. Overall， the proposed method can learn discriminative crop growth features without using labels during training and can serve as a preliminary tool for pre-labeling analysis and data organization in large-scale aerial agricultural imagery.
