水稻是台灣重要的作物之一，每年政府都會需要了解其種植的區域與面積，並用於統計產量及訂定相關決策。傳統方法是經由專家對每張遙測影像進行判釋並人工繪製標記，然而這樣的方式效率很低，根本無法及時的提供大量調查的資訊。隨著近年人工智慧技術的發展，把相關技術應用於輔助判釋便可以大幅降低該項作業對人力的依賴，並能做到快速有效的水稻自動判釋。 目前應用於水稻判釋的方法是讓模型學習整張航照影像的資訊，然而對模型來說可能需要學習的資訊太多，因此本研究提出讓模型只在農地上作學習。另外，目前採用的語義分割模型是UNet，因其缺乏考慮像素間的空間關係，導致分割結果呈現破碎狀態，造成準確率降低，本研究將提出UNet-FNN架構來解決此問題。根據觀察，在不同地區的稻田有許多不同的樣貌，因此本研究將使用基於圖幅塊的隨機採樣來增加訓練資料的多樣性以訓練出更好的模型。;Rice is one of the important crops in Taiwan. Every year, the government needs to know its planting area and the location, and use it to calculate the yield and make relevant decisions. The traditional method is that experts interpret each remote sensing image and draw labels manually. However, this method is very inefficient and cannot provide a large amount of survey information in time. With the development of artificial intelligence technology in recent years, the application of related technologies to assisted interpretation can greatly reduce the dependence on manpower, and achieve rapid and effective automatic interpretation of rice. The current method used in rice segmentation is that model learns the information of the entire aerial image. However, there may be too much information for the model to learn. Therefore, this research proposes that the model only learns on farmland. In addition, the currently used semantic segmentation model is UNet. Because UNet does not consider the spatial relationship between pixels, it will lead to speckle segmentation results, resulting in reduced accuracy. This research will propose the UNet-FNN architecture to solve this problem. According to observations, there are many different appearances of rice fields in different area. Therefore, this study will use random sampling based on framelets to increase the diversity of training data to train a better model.
