2025-03-14
【學術亮點-頂級期刊論文】應用領域適應人工智慧提升多期作物之穀粒含水量預測
字體大小
小
中
大
【學術亮點-頂級期刊論文】應用領域適應人工智慧提升多期作物之穀粒含水量預測
Intelligent Cultivation: Crop digital twin for growth adversity index and disaster damage assessment【Department of Civil Engineering / Ming-Der Yang / Tenured Distinguished Professor】
智慧栽培：數位分身作物生長逆境指標與災損監測【土木工程學系楊明德終身特聘教授】
上架日期2025-04-01
Intelligent Cultivation: Crop digital twin for growth adversity index and disaster damage assessment【Department of Civil Engineering / Ming-Der Yang / Tenured Distinguished Professor】
智慧栽培：數位分身作物生長逆境指標與災損監測【土木工程學系楊明德終身特聘教授】
| 論文篇名 | 英文：Enhancing grain moisture prediction in multiple crop seasons using domain adaptation AI 中文：應用領域適應人工智慧提升多期作物之穀粒含水量預測 |
| 期刊名稱 | Computers and Electronics in Agriculture (指標清單期刊) |
| 發表年份, 卷數, 起迄頁數 | 2025, 231, no.110058 |
| 作者 | Yang, Ming-Der (楊明德); Hsu, Yu-Chun*; Liu, Tsai-Ting; Huang, Han-Hui |
| DOI | 10.1016/j.compag.2025.110058 |
| 中文摘要 | 在精準農業中，基於影像的人工智慧模型之實際性能通常會受目標域和源域之差異，如照明、光譜反射率、物種、生長階段和作物季節等變化的不利影響。本研究開發一機器學習領域自適應 (DA) 工作流程，以量化稻米的穀物水分含量 (GMC) 並增強模型在不同作物季節和條件的穩健性。使用行動裝置影像和涵蓋 9 個生長季節（2019–2023 年）、38 個物種和 2 個背景條件的 GMC 調查資料建立 GMC 資料集。使用U2 - Net和YOLOv7模型對影像進行預處理，提取和校正穗的幾何和光譜特徵。評估 11 種機器學習模型和 22 種 DA 演算法，以確定 GMC 預測的最佳模型。使用基於特徵、實例和參數的演算法進行監督和非監督 DA 實驗。基於實例的演算法表現了最低平均絕對誤差（MAE）--0.28％。目標域數據顯示，具真實數據比率為 0.1% 的監督 DA 具有較高性能，MAE 為 1.51%。當 GMC 低於 40% 時，MAE 為 1.07%，比無 DA 的模型提高 54.66%，最佳收成時間之誤差為 ±1 天。研究結果在 900 公頃稻田上進行測試，GMC 間隔差異在 1.2% 內，可證明模型穩健性。此 DA 工作流程可增強農業模型適應性，在精準農業中具有極高潛力。 |
| 英文摘要 | In precision agriculture, the real-world performance of image-based artificial intelligence models is often adversely affected by variations in, for example, illumination, spectral reflectance, species, growth stages, and crop seasons between the target and source domains. In this study, a machine learning domain adaptation (DA) workflow was developed to quantify the grain moisture content (GMC) of rice and enhance model robustness across crop seasons and conditions. A GMC data set was established using imagery from mobile devices and GMC survey data covering 9 growing seasons (2019–2023), 38 species, and 2 background conditions. U2-Net and YOLOv7 models were used to preprocess the imagery and extract and correct panicles’ geometric and spectral features. A total of 11 machine learning models and 22 DA algorithms were evaluated to identify the optimal model for GMC prediction. Experiments with supervised and unsupervised DA were conducted using feature-, instance-, and parameter-based algorithms. The instance-based algorithms achieved the lowest mean absolute error (MAE), 0.28 %. The target domain data revealed that supervised DA at a 0.1 % data ratio has moderated high performance, with a MAE of 1.51 %. When the GMC was below 40 %, the MAE was 1.07 %, indicating a 54.66 % improvement over models without DA, with an error of ±1 day. These results were visualized over a 900-ha rice field, demonstrating the models’ robustness, with GMC interval differences within 1.2 %. This DA workflow can enhance model adaptability in agriculture, indicating it holds potential in precision agriculture. |
| 發表成果與AI計畫研究主題相關性 | 利用領域適應人工智慧模型辨識跨期作之手機照片中穀粒含水量，提高模型準確度54.66%，所建議之最佳收成時間之誤差為 ±1 天。 |
