A novel method for synchronous retrieval of land surface temperature and emissivity based on DLMSR-Transformer-MoE model

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.