彩色紅外線(英語：Color-Infrared, 縮寫 CIR)航照圖幅中的水稻坵塊識別可揭示農作物種植園的位置和面積等資訊，從而協助政府制定政策。先前針對水稻坵塊識別的研究[1]已經為此任務實現了一個封閉集合分類器，並且它被證明在推理論過程中容易受到新穎坵塊的影響而產生誤判。該封閉集合分類器由於缺乏分布外檢測，在推論階段是不穩健的。本文提出了一種基於開放集合識別分類重建學習 (英語：Classification Reconstruction Learning for Open Set Recognition，縮寫CROSR) 的新穎坵塊檢測方法。具體來說，我們用名為 RecUNet 的特製深度神經網絡替換了原 CROSR 架構中用於提取特徵向量的神經網絡 DHRNet，RecUNet 有助於保存低階語義密集特徵圖。實驗證明 RecUNet 相較於 DHRNet 可以在 AUROC 效能評估指標帶來百分之十五的提升。此外，我們提供的初步結果顯示在航照圖幅內檢測到新穎坵塊時，使用每個航照圖幅中的新穎坵塊檢測率來預測逐圖幅水稻識別性能下降是可行的。;Rice parcel recognition in Color Infrared (CIR) aerial images reveals information such as location and area of crop plantations that assist government policymaking. Prior work on rice parcel recognition[1] has implemented a closed-set classifier for this task, and it was shown to be vulnerable to novel parcels during inference. Without outlier a detection mechanism, the close set classifier is not robust during inference. This paper presents an approach for novel parcel detection in CIR aerial image frames based on Classification-Reconstruction learning for Open-Set Recognition (CROSR). Specifically, we replaced the neural network for latent vector extraction DHRNet with a customized deep neural network named RecUNet that facilitates preservation of low-level semantic dense representation maps. Experimental validation demonstrated that RecUNet brought an improvement of 15 percent in terms of AUROC compared to DHRNet. Furthermore, we provide preliminary findings showing that upon detection of novel parcels within aerial image frames, it is feasible to predict frame-wise rice recognition performance drops using per aerial image frame novel parcel detection ratios.
