航空攝影技術廣泛應用於農業調查，結合深度學習技術，可以解析高 精度的航空照片內容，協助調查人員進行農地面積計算和農作坵塊的農作 物分類，進而瞭解農業土地利用現況，節省人力和調查時間。然而，訓練 資料的收集通常非常耗時，且目前的資料集是在多年的時間裡逐一收集而 成。在這樣的背景下，目前已有多模型架構的方法來解決這個問題，但這 些方法需要大量的推論時間。在實際應用中，使用者往往無法接受如此長 的推論時間。因此，本研究引入增量式學習的方法來解決推論時間過長的 問題。增量式學習允許模型在不重新訓練的情況下，接受新的訓練資料並 進行更新，從而節省了推論時間。通過增量式學習，我們可以有效地利用 已有的模型知識，並隨著新的資料進行模型更新，以適應新的農作物類型 和土地利用情況。透過引入增量式學習的方法，我們可以在不犧牲準確性 的情況下，實現更快速的推論時間，從而使航空攝影技術在農業調查中更 具實用性和效率。;Aerial photography technology is widely used in agricultural surveys, and when combined with deep learning techniques, it can provide detailed interpreta tion of high-resolution aerial images. This assists surveyors in calculating agri cultural land areas and classifying crops in agricultural parcels, thereby gaining insights into agricultural land use patterns and saving manpower and survey time. However, the collection of training data for such applications is time-consuming and often spans multiple years. In this context, existing approaches have employed multi-model architectures to address the problem. However, these methods require significant inference time, which is not acceptable for practical use by end-users. To overcome the issue of prolonged inference time, this study introduces an incre mental machine learning approach. Incremental learning allows the model to ac cept new training data and update itself without retraining, thereby reducing infer ence time. By leveraging incremental learning, we can effectively utilize existing model knowledge and update the model with new data to adapt to new crop types and land use conditions. By incorporating incremental learning, we can achieve faster inference time without compromising accuracy, making aerial photography technology more practical and efficient in agricultural surveys.
