Enhanced 10m Remote Sensing Ecological Index (RSEI) and Land Surface Temperature (LST) over Taiwan in 2023 Based on Integreted Deep Learning Super-resolution Framework

This dataset presents the results of an enhanced 10 m Remote Sensing Ecological Index (RSEI) over Taiwan, derived using super-resolved Sentinel-2 imagery based on the proposed Sen2-SSIR model. The dataset aims to provide a high-resolution ecological assessment to support environmental monitoring and analysis. In addition, the dataset includes Land Surface Temperature (LST) downscaling results at a 10 m spatial resolution. The downscaling process was conducted using a DNNr (Deep Neural Network regression) model, which was subsequently utilized in the construction of the final enhanced RSEI. Abstract :Ecological quality assessment based on remote sensing data typically depends on satellite sensors equipped with thermal infrared bands to derive land surface temperature (LST). However, these sensors generally provide spatial resolutions of approximately 30m or coarser, limiting their suitability for fine-scale ecological monitoring. Although Sentinel-2 offers finer spatial detail and rich multispectral information, it lacks thermal bands required to construct remote sensing-based ecological analysis. Moreover, Sentinel-2 consists of 13 spectral bands of which only four are available at 10m spatial resolution, while the remaining key bands are provided at 20m which limits the 10m resolution ecological analysis. To address these limitations, this study proposes an integrated deep learning framework to construct enhanced 10m remote sensing-based ecological analysis in Taiwan. First, a novel Sentinel-2 super-resolution based on deep learning model (termed Sen2-SSIR) is developed to enhance all 20m bands to 10m while preserving spectral consistency. Sen2-SSIR model integrates Inception module for multi-scale feature extraction, Residual-in-Residual Dense Block (RRDB) module for deep feature learning by integrating multi-level residual networks with dense connections, and post-residual reconstruction module for spatial-spectral detail recovery. Second, addressing the absence of thermal bands in Sentinel-2, deep neural network regression model was used to generate 10m LST. Finally, super-resolved Sentinel-2 and downscaled LST are integrated to construct enhanced remote sensing-based ecological data at 10m spatial resolution. Results show that Sen2-SSIR outperforms conventional resampling and existing deep learning approaches. In addition, enhanced 10m remote sensing-based ecological analysis provides more spatially detailed assessment and indicating Taiwan exhibits moderate ecological condition. The source code and supplementary datasets associated with this study are available upon request by email to [e-mail]本数据集呈现了基于所提出的Sen2-SSIR模型，利用超分辨率处理后的Sentinel-2影像生成的台湾地区增强型10米分辨率遥感生态指数（Remote Sensing Ecological Index, RSEI）结果。本数据集旨在提供高分辨率生态评估数据，以支撑环境监测与分析工作。 此外，本数据集还包含10米空间分辨率的地表温度（Land Surface Temperature, LST）降尺度结果。该降尺度过程采用深度神经网络回归（Deep Neural Network regression, DNNr）模型完成，后续该模型被用于构建最终的增强型RSEI。 摘要：基于遥感数据的生态质量评估通常依赖搭载热红外波段的卫星传感器来反演地表温度（Land Surface Temperature, LST），但此类传感器的空间分辨率通常仅约30米或更低，限制了其在精细尺度生态监测中的应用。尽管Sentinel-2具备更精细的空间细节与丰富的多光谱信息，但其缺少构建遥感生态分析所需的热红外波段。此外，Sentinel-2共包含13个光谱波段，其中仅4个波段具备10米空间分辨率，其余关键波段的分辨率均为20米，这制约了10米分辨率的生态分析工作。为解决上述局限，本研究提出了一套集成深度学习框架，以构建台湾地区10米分辨率的增强型遥感生态分析产品。首先，本研究开发了一种基于深度学习的新型Sentinel-2超分辨率模型（命名为Sen2-SSIR），可在保持光谱一致性的前提下，将所有20米分辨率波段提升至10米。Sen2-SSIR模型整合了用于多尺度特征提取的Inception模块、通过融合多级残差网络与密集连接实现深度特征学习的残差密集块（Residual-in-Residual Dense Block, RRDB）模块，以及用于空间光谱细节恢复的后残差重构模块。其次，针对Sentinel-2缺失热红外波段的问题，本研究采用深度神经网络回归模型生成10米分辨率的LST数据。最后，将超分辨率处理后的Sentinel-2影像与降尺度后的LST数据进行融合，构建得到10米分辨率的增强型遥感生态数据。实验结果表明，Sen2-SSIR模型的性能优于传统重采样方法与现有深度学习方法。此外，增强型10米分辨率遥感生态分析可提供更精细的空间评估结果，显示台湾地区整体生态状况处于中等水平。 本研究相关的源代码与补充数据集可通过发送邮件至[e-mail]申请获取。