Sentinel-2 因其免費存取、豐富的光譜資訊和良好的時空覆蓋範圍而被廣泛應用於遙感應用。 Sentinel-2 由 13 個光譜波段組成，其中只有 4 個波段的分辨率為 10 米，其餘波段的分辨率較低，為 20 米和 60 米。然而，10m 和 20m 光譜波段通常用於土地和植被分析，包括生態研究。本研究提出基於深度學習模型的Sentinel-2單張影像超解析度（命名為Sen2-SSIR），將20m波段的空間解析度提升至10m。 Sen2-SSIR 由三個模組組成，包括 inception 模組、residual-in-residual 密集塊 (RRDB) 模組和 post-residual 模組。 Inception 模組將用作初始學習過程，以有效地學習來自精細和粗分辨率帶的連接特徵圖。 RRDB 模組旨在透過整合具有密集連接的多層殘差網路進行深度特徵學習。同時，利用後殘差模組從初始和深度特徵圖中恢復空間和光譜細節。 Sen2-SSIR 有效地學習了高解析度波段的空間細節，同時保留了粗分辨率波段的光譜資訊。研究位於台灣，使用台灣境內的Sentinel-2場景作為訓練和測試資料。所提模型的 RMSE、SRE 和 ERGAS 分數（影像品質指標）分別為 58.61、88.11 和 0.55。這些結果表明 Sen2-SSIR 優於其他傳統影像重採樣方法和其他基於深度學習的超解析度模型。下一階段，我們將利用超分辨的10顆Sentinel-2衛星的結果，進行台灣地區的RSEI分析。調查結果顯示，台灣生態狀況屬中等，平均值為0.69。;Sentinel-2 is widely used in remote sensing applications due to its free accessibility, rich spectral information, and fine spatial-temporal coverage. Sentinel-2 consists of 13 spectral bands with only four bands at 10m resolution, while the remaining bands have lower resolutions of 20m and 60m. However, the 10m and 20m spectral bands are commonly used for land and vegetation analysis including ecological studies. This research proposed single image super-resolution of Sentinel-2 based on deep learning model (named Sen2-SSIR) to enhance the spatial resolution of 20m bands to 10m. Sen2-SSIR consists of three modules, including inception module, residual-in-residual dense block (RRDB) module, and post-residual module. Inception module will be used as initial learning process to effectively learn concatenated feature maps from fine and coarse resolution bands. RRDB module is designed for deep feature learning by integrating multi-level residual networks with dense connections. Meanwhile, post-residual module is used to recover the spatial and spectral details from the initial and deep feature maps. The Sen2-SSIR effectively learns the spatial details from fine-resolution bands while preserving the spectral information of coarse-resolution bands. The study is located in Taiwan and uses the Sentinel-2 scenes within Taiwan as the training and testing data. The RMSE, SRE, and ERGAS scores (image quality metrics) from proposed model are 58.61, 88.11, and 0.55, respectively. These results demonstrate that Sen2-SSIR outperforms other conventional image resampling methods and other deep learning-based super-resolution models. The results of the super-resolved 10 Sentinel-2 were used for RSEI analysis in Taiwan. It was found that Taiwan has a moderate ecological status with an average value of 0.69.
