| 摘要: | 隨著農業自動化與智慧化的發展，如何有效掌握作物生長情形成為農政單位提升農業管理效率之關鍵課題。水稻為臺灣最重要的糧食作物之一，目前水稻生長監控主要仰賴 1/5000 航照正射影像及水稻專家進行人工標記，雖具高度準確性，然此方式不僅耗時且需大量人力，限制其於大規模應用之可行性。
為了降低對人工標記資料的依賴，本研究採用對比學習技術，提出一套針對水稻坵塊影像之自動分群系統，旨在無需人工標記資料的前提下，精確判斷水稻坵塊所處生長階段，進而實現自動化的作物分群與標記流程。核心技術採用穩定群集辨別（Stable Cluster Discrimination, SeCu）[1] 技術，透過移除負樣本梯度、改良交叉熵損失函數，使模型於無監督訓練下亦能學習穩定且具語意之影像特徵表示。模型完成訓練後，不需搭配分群演算法即可直接針對水稻坵塊進行分群，達成水稻子類別（插秧期、生長期、黃熟期、收割期）的自動分群。
本研究以臺灣實際航照影像與農業地理資訊為資料來源，建置水稻坵塊影像資料集，並針對不同方法進行實驗比較，採用分群準確率（Clustering Accuracy, ACC）、調整蘭德指數（Adjusted Rand Index, ARI）與正規化互資訊（Normalized Mutual Information, NMI）作為評估指標。實驗結果顯示，本研究所提出之方法相較於現有方法效果皆有明顯提升，分別達成 ACC: 0.8500、ARI: 0.6927 與 NMI: 0.7096。
綜上所述，本研究不僅提供一種高效且低成本的水稻子類別分群解決方案，亦驗證 SeCu 方法於農業影像應用上的潛力，對於未來智慧農業發展具有高度參考價值與實務應用前景。;With the advancement of agricultural automation and intelligent technologies, effectively monitoring crop growth has become a key issue for government agencies to improve agricultural management efficiency. As one of the most important staple crops in Taiwan, rice growth monitoring currently relies heavily on 1/5000 aerial images and manual annotation by experts. Although this approach yields high accuracy, it is both time-consuming and labor-intensive, limiting its scalability for large-scale applications.
To reduce reliance on manually labeled data, this study adopts contrastive learning techniques and proposes an automatic clustering system for rice parcel images. The system aims to identify the growth stage of rice parcels without requiring human annotations, thereby enabling automated crop categorization and labeling. The core of the system leverages the Stable Cluster Discrimination (SeCu) method [1], which enhances feature stability and semantic representation by removing negative sample gradients and modifying the cross-entropy loss function. Once training is completed, the model can directly cluster rice parcels without the need for additional clustering algorithms, effectively distinguishing the four rice growth stages: transplanting, growing, ripening, and harvesting.
This study constructs a rice parcel image dataset based on real aerial imagery and agricultural geographic data from Taiwan. Comparative experiments are conducted using multiple methods, with Clustering Accuracy (ACC), Adjusted Rand Index (ARI), and Normalized Mutual Information (NMI) as evaluation metrics. The results show that the proposed method significantly outperforms existing approaches, achieving ACC: 0.8500, ARI: 0.6927, and NMI: 0.7096.
In summary, this research provides an efficient and low-cost solution for rice growth stage clustering, demonstrating the potential of the SeCu method in agricultural image applications and offering valuable insights and practical prospects for future smart agriculture development. |
