A novel method for synchronous retrieval of land surface temperature and emissivity based on DLMSR-Transformer-MoE model
Description
This dataset supports the study “DLMSR–Transformer–MoE: A novel method for synchronous retrieval of land surface temperature and emissivity”. The dataset was constructed under the assumption that multi-channel thermal infrared brightness temperatures can be used to learn the nonlinear relationship between LST and LSE when physical solvability constraints, product-level refinement, and surface-adaptive modeling are incorporated. The data include preprocessed Aqua/MODIS thermal infrared observations and auxiliary products obtained from NASA Earthdata, including brightness temperatures, geolocation information, cloud masks, and original LST&E products. These data were processed through quality control, cloud screening, geometric correction, spatial resampling, mosaicking, variable extraction, and conversion into text-format files for model training and validation. The dataset can be used to reproduce the DLMSR–Transformer–MoE retrieval workflow, including model training, global multi-temporal cross-validation, in situ LST validation, and classification-based LSE validation. Brightness temperature variables should be interpreted as model inputs, while original and refined LST&E values provide supervisory or reference information. The validation data can be used to evaluate the consistency between the model retrievals, MODIS products, ground-based observations, and classification-based emissivity references. Users can reproduce the study by using the processed files provided here or by downloading the same Aqua/MODIS products from NASA Earthdata and applying the preprocessing and model codes described in the associated manuscript.
Files
Steps to reproduce
The remote sensing data used in this study were mainly obtained from the Aqua/MODIS products provided by the National Aeronautics and Space Administration (NASA). According to the objectives of this study, the required MODIS thermal infrared observations and auxiliary products were first downloaded from the NASA Earthdata platform, including brightness temperature data, geolocation information, cloud mask data, and the original land surface temperature and land surface emissivity products. The raw remote sensing data were then processed through a standardized preprocessing workflow, including data reading, quality control, cloud detection and removal, geometric correction, spatial resampling, projection transformation, and global-scale mosaicking. These steps were conducted to ensure spatial, temporal, and resolution consistency among different MODIS products. After preprocessing, the input variables and target variables required by the retrieval model were extracted, and the raster data were converted into text-format datasets for model training, cross-validation, and accuracy assessment. To facilitate reproducibility, the data-processing procedure in this study follows a unified preprocessing and model-execution workflow. Researchers seeking to replicate this work can download the same Aqua/MODIS products from the NASA Earthdata platform and process them following the steps described above, including geometric correction, resampling, mosaicking, quality control, and variable extraction. Alternatively, they may directly use the processed datasets provided in this study, convert them into the text format required by the model, and run the corresponding codes for model training, prediction, and validation. No specialized reagents or laboratory instruments were required in this study. The workflow mainly relies on publicly available remote sensing data, standard remote sensing preprocessing procedures, and the deep learning retrieval codes developed in this work. Therefore, the results can be reproduced by obtaining the same data sources, applying the same preprocessing procedures, and executing the same model codes.
Institutions
- Chinese Academy of Agricultural SciencesBeijing, Beijing