以多時序SAR衛星資料建立臺灣地區水稻坵塊深度學 習判釋模型

Abstract

長期以來，遙測影像分析技術可提供快速且大面積的地表覆蓋現況資訊，提供政府機關、學術單位與民間團體進一步分析土地利用的發展歷程，作為土地合宜利用規劃與成長管理的重要資訊來源。在各類型土地覆蓋樣態當中，以農業類型土地利用的變動樣態最為多元、快速、易受自然環境事件及人為事件與喜好而影響分布區位，但卻又多有穩定的循環週期規則，往往造成難以估算種植面積、生產狀態、收穫產量及市場價格變化。 為強化農業土地利用的監控分析成果，縮短資料整合與分析時間成本，提升整體管理效益，本研究從空間資訊整合角度，基於臺灣地區現有的農地管理坵塊資料、多元與多時序航遙測空間資訊萃取機制，提出可整合空間向量、地表時序變化特徵與現地調查及管理清冊資料的多維度土地覆蓋特徵描述資料架構，並以深度學習模型中的雙向長短期記憶法（Bi-Directional Long-Short Term Memory，Bi-LSTM）設計判釋模型架構，發展適用於農業土地利用集約且具農地管理空間資料的多維度空間資訊水稻坵塊深度學習判釋模型。 本研究分別以雲林與嘉義地區於2021年二期稻作與2022年一期稻作的多時期Sentinel-1雷達影像後向散射係數資料為遙測資料基礎，透過本研究所建立之特徵描述資料架構與Bi-LSTM進行水稻坵塊判釋，於2021年二期稻作與2022年一期稻作資料組合中，達成87.70~94.66% （Kappa 0.75~0.89）與85.03~91.58% （Kappa 0.67~0.78）的整體判釋正確性。透過本研究的實證成果，雖受限於影像資料空間解析度僅達10m、光學衛星影像多受雲覆影響而未能有效整合、Sentinel-1衛星受太陽風影響資料接收及Sentinel-1B衛星於2022年失效等條件，但目標區域農地坵塊的整體判釋正確性仍能達到85%以上的成果。 本研究所提出的多維度空間資訊水稻坵塊深度學習判釋模型，是以描述土地使用的空間單元為基礎，藉由多元多時序的遙測影像資料建立空間單元在時序上與多元影像資料間的特徵關係，再透過深度學習模型進行判釋。因此在本研究所提出的模型架構上，可配合空間資料蒐整狀況與深度學習模型的發展，彈性調整與擴充本研究模型的空間單元數量、多元多時序遙測資料及深度學習判釋模型，藉以符合多樣的土地覆蓋現況判釋與監測需求，進而達成快速且有效偵測農地利用現況資訊的目標。

For a long time, remote sensing image analysis technologies can provide rapid and extensive land cover information for government agencies, academic institutions, and private organizations to planning appropriate land policies or usage management. Among various land cover patterns, agricultural land covers exhibit diverse, rapid, and easily influenced variations due to natural environmental events, human activities, and owner’s preferences. But most of them often follows stable cyclic patterns. Those characteristics of agriculture land cover cause challenging to estimate cultivation area, productions, and market price variance. To enhance the monitoring and analysis of agricultural land use, reduce data integration and analysis costs, and improve management efficiency, this study proposes a multidimensional land cover feature description data framework based on spatial information integration and demonstrated the integration of deep learning interpretation model. This study incorporates multidimensional spatial information, including land cover temporal change features, survey and management inventory data. Additionally, a deep learning model based on Bidirectional Long-Short Term Memory (Bi-LSTM) is designed for developing an appropriate multidimensional spatial information rice field’s interpretation model via deep learning for intensive agriculture land usage area. The study applies multi-temporal Sentinel-1 SAR data for the 2nd rice season in 2021 and the 1st rice season in 2022 in the Yunlin and Chiayi counties. Using the multidimensional information description framework and Bi-LSTM, the study achieves overall interpretation accuracies ranging from 87.70% to 94.66% (Kappa 0.75 to 0.89) for the 2nd rice season in 2021, and 85.03% to 91.58% (Kappa 0.67 to 0.78) for the 1st rice season in 2022. Despite limitations such as the 10m spatial resolution of SAR data, interference from cloud cover in optical satellite imagery, data reception issues with Sentinel-1 due to solar wind interference, and the failure of Sentinel-1B satellite in 2022, the overall interpretation accuracy for the targeted agricultural land fields still exceeds 85%. The multidimensional spatial information rice field’s deep learning interpretation model proposed in this study is based on the temporal and feature relationships between spatial units and diverse multi-temporal remote sensing data, and implemented deep learning model to interpret land use. Therefore, the model's architecture proposed in this study can be flexibly adjusted and expanded based on the spatial data collection and the development of deep learning models. This enables the model to adapt to diverse land cover interpretation and monitoring requirements, achieving the goal of rapidly and effectively detecting information on current agricultural land use.