水稻為臺灣最重要的糧食作物之一，主管機關須及時掌握其種植分布與插秧時程， 以支援糧食安全、糧價穩定與轉作輔導等政策決策。Sentinel-2 衛星提供高時序且免費 的多光譜影像，為大範圍農情監測帶來契機，然而臺灣場景具有亞熱帶春季頻繁雲遮、 小農制坵塊破碎，以及年度間物候漂移等難點；既有的臺灣水稻遙測研究多以分類製 圖為目標，水稻物候日期亦多作為製圖的中間產物，缺乏將分類與日期提取相互串接 的整合框架。 為回應農業政策對「在哪裡種」與「何時種」的雙重需求，本研究以坵塊化的Sentinel-2 時序為輸入，針對臺灣一期作水稻建構一套可於生長季中運作的兩階段整 合框架。Stage 1 以五個深度學習模型進行 soft voting 集成，於坵塊層級判釋水稻／非 水稻；Stage 2 則對判定為水稻的坵塊，以常態化差值植生指標（Normalized Difference Vegetation Index， NDVI）與修正型常態化差值水體指標（Modified Normalized Difference Water Index， MNDWI）差值序列的谷底偵測插秧期的淹水物候，據以估計插秧日，即 該坵塊之生長季開始（Start of Season， SOS），並為每筆估計附加物理可解釋的品質標 記。 實驗結果顯示，Stage 1 具早期可用性，水稻類 F1 於 3 月底即達約 0.91、並於 4 月 後趨於穩定，跨年度測試下效能的下降幅度有限。Stage 2 已套用於全臺 744，938 筆水稻 坵塊，其插秧日之估計經三層獨立證據檢驗：街景地面真值的平均絕對誤差約 18 天、 以抽穗為錨點之模板匹配的中位絕對偏差約 5 天，且插秧 V 形特徵於跨年度間具一致 性；以局部極小值定位谷底的設計，亦使偵測能隨觀測累積而支援季中的滾動更新。 本研究將早期水稻分類與插秧日提取整合為一套自動化、可早期化且可逐年複現的可 部署監測框架，並透過特徵消融與兩階段整合分析界定其能力邊界，為臺灣水稻的季 中農情掌握提供了一條低成本而可操作的途徑。;Rice is one of Taiwan’s most important staple crops， and the authorities require timely knowledge of its spatial distribution and transplanting schedule to support policy decisions such as food security， price stabilization， and crop-rotation guidance. Sentinel-2 provides hightemporal-frequency and freely available multispectral imagery， creating an opportunity for largescale agricultural monitoring. The Taiwanese setting， however， poses several difficulties: frequent and prolonged cloud cover during the subtropical spring， highly fragmented parcels under small-holder farming， and inter-annual phenological drift. Existing rice remote-sensing studies in Taiwan have largely targeted classification mapping and have treated rice phenological dates as intermediate products， lacking an integrated framework that links classification with date extraction. To address agricultural policy’s dual need to know ”where rice is grown” and ”when it is transplanted，” this study develops a two-stage integrated framework that operates within the growing season， taking parcel-level Sentinel-2 time series as input， for first-crop rice in Taiwan. Stage 1 performs parcel-level rice/non-rice classification through a soft-voting ensemble of five deep-learning models. Stage 2， applied only to parcels classified as rice， detects the flooding phenology of transplanting from the valley of the NDVI–MNDWI difference series to estimate the transplanting date—that is， the parcel’s Start of Season (SOS)， attaching a physically interpretable quality flag to each estimate. The results show that Stage 1 offers early-season usability: the rice-class F1 reaches approximately 0.91 by late March and stabilizes after April， with limited performance degradation under a cross-year setting. Stage 2 has been applied to 744，938 rice parcels across Taiwan， and its transplanting-date estimates were examined through three independent lines of evidence: a mean absolute error of about 18 days against street-view ground truth， a median absolute deviation of about 5 days against heading-anchored template matching， and consistency of the transplanting V-shape across years. The use of local-minimum (argmin) valley localization further enables in-season rolling updates as observations accumulate. By integrating early rice classification with transplanting-date extraction into an automated， early-available， and annually reproducible operational monitoring framework， and by delineating its capability boundaries through feature ablation and two-stage integration analysis， this study offers a low-cost and operable approach to in-season agricultural monitoring of rice in Taiwan.
