Research on anomaly detection and correction method for remote sensing instrument calibration data based on TELSTM-ADC model

Radiometric calibration is fundamental to quantitative remote sensing. The Microwave Radiation Imager (MWRI), one of the key instruments onboard the Fengyun-3D (FY-3D) satellite, requires accurate hot load reflector temperature measurements to ensure reliable brightness temperature computation. After the FY-3D satellite had been on orbit for five years, the hot load temperature measurement module began to suffer from intermittent failures caused by the harsh space environment, making it impossible to obtain valid temperature data. To address this, we developed a Two-stage Enhanced Long Short-Term Memory for Anomaly Detection and Correction (TELSTM-ADC). The method streamlines the LSTM architecture by merging the input and forget gates into a single update gate, integrates a Bayesian network for adaptive parameter updates to capture time-varying behaviour, and applies segmented iterative forecasting for stable long-sequence predictions. Validation with independent nominal operational data shows that RMSE = 0.62 K, MAE = 0.61 K, and STD = 0.14 K for temperatures ranging from 200K to 400K. Test results for the failure-period hot load reflector temperature data show that the TELSTM-ADC model can effectively detect and correct anomalies, with the corrected temperatures accurately reproducing historical variation patterns. When the hot load reflector temperature data are applied to radiometric calibration and cross-compared with GPM/GMI brightness temperatures (used as a radiometric reference), the results demonstrate that the TELSTM-ADC – based predictive correction restores the quality of the observed brightness temperatures to a level essentially consistent with that prior to the system failure.