運用卷積長短期記憶網路建立土地覆蓋變遷推估模式與探究

Abstract

都市土地的使用是引發都市環境變遷的重要驅動力。運用推估模式的建置持續監測都市土地變遷，得以增加環境衝擊時的因應能力。卷積長短期記憶網路ConvLSTM具有對時空特徵的存儲與記憶能力，適合對未來的空間配置進行推估。考慮到ConvLSTM模型架構在土地使用變遷領域中對土地使用變遷的推估尚未對模型有著正式的適合性探討，本研究以遙測衛星影像萃取出多時期都市土地覆蓋類別作為研究資料，設計並訓練出一套推估都市土地覆蓋變遷的卷積長短期記憶網路模型，同時萃取不同衛星資源的衛星影像對訓練好的模型進行驗證，評估模型是否具有良好的穩健特性。最後預測對2022上半年度的土地覆蓋推估整體準確率為74%，認定該模型適合應用於建置都市土地覆蓋變遷推估模式上，期待給予後續研究一定的啟發，並能在往後透過土地覆蓋變遷推估模式的建置對都市地區進行輔助模擬與監測，達到都市化發展中決策輔助的效果。Urban land use is an important driving force for urban environmental changes. Using the prediction model to continuously monitor urban land changes can increase the ability to respond to environmental shocks. The Convolutional Long Short-Term Memory network has the ability to store and memorize spatiotemporal features, suitable for predicting future spatial pattern. Considering that the ConvLSTM model architecture has not yet discussed the suitability of the model for the prediction of land use change in the field of land use change formally, this study uses satellite images to extract multi-period urban land cover categories as research data. A set of ConvLSTM network models for predicting urban land cover changes. At the same time, satellite images of different satellite resources are extracted to verify the trained model and evaluate whether the model has good robust characteristics. Finally, it is predicted that the overall accuracy rate of land cover estimation in the first half of 2022 is 74%, and it is determined that the model is suitable for the establishment of urban land cover change prediction models. which can assist in the simulation and monitoring of urban areas, so as to achieve the effect of decision-making assistance in urbanization development.