Automated vicarious radiometric validation of spaceborne thermal infrared sensors at non-dedicated validation sites using deep learning-based cloud filtering

• Automated validation method for thermal infrared remote sensing is proposed. • Non-dedicated in-situ temperatures at lakes and bays are used as ground truth. • The Swin Transformer classifier is applied to filter out cloudy image. • Automatically found large number of matchups between satellite and in-situ data. • The study sites were adequate for validation of thermal infrared remote sensing. Vicarious calibration/validation (Cal/Val) is essential for ensuring the radiometric accuracy of spaceborne thermal infrared (TIR) sensors. Anticipating an increase in the number of TIR sensors in the near future, this study developed an automated vicarious radiometric validation method based on in situ water temperature measurements from 14 sites in lakes and bays across Japan, which are not specifically dedicated to Cal/Val of satellite sensors. In addition, we applied a contextual image classification model based on the Swin Transformer architecture to create a fully automated filtering procedure to remove cloudy data. The proposed methods were developed and evaluated using the data acquired by well-calibrated satellite sensors, namely Advanced Spaceborne Thermal Emission and Reflection Radiometer Thermal Infrared Radiometer (ASTER TIR), Landsat 8 Thermal Infrared Sensor (TIRS), and Landsat 9 TIRS-2, as reference targets. To develop a contextual image classification model, we fine-tuned a pre-trained Swin Transformer model with our own training data comprising 921,967 chip images created from ASTER TIR band 13 and Landsat 8 TIRS band 10 images selected from global regions. The developed image classification model achieved an overall accuracy of 96.20 %. However, when applied only to the study areas, the accuracy decreased to 94.85 %, because all the target sites were exclusively located in lakes and bays. The classification model was localized to our study areas by adjusting the probability threshold. Combining contextual image classification with quantitative thresholds, the model successfully classified 90 % of the cloud-free daytime ASTER data and Landsat 8 data. The accuracy for cloud-free classification was 81 % and 86 % for nighttime ASTER data and Landsat 9 data, respectively. Consequently, 169, 58, 500, and 130 matchups were automatically identified for daytime ASTER, nighttime ASTER, Landsat 8, and Landsat 9, respectively. The in situ water temperature for each matchup was converted to top-of-atmosphere brightness temperature (TOA BT) through radiative transfer calculations. In situ -based and satellite-based TOA BT agreed very well within the residual bias error of less than ±0.4 K except for nighttime ASTER data that seemed to be affected by insufficient skin temperature correction. The correlation between in situ -based and satellite-based TOA BT was strong, with R 2 values ranging from 0.97 to 0.99, for daytime and nighttime ASTER, Landsat 8, and Landsat 9. The statistically estimated offsets between the in situ -based and satellite-based TOA BT were nearly equivalent to previously reported Cal/Val result and within the acceptable range of the sensors’ requirements, indicating that our method and data are suitable for the radiometric validation of spaceborne TIR sensors.