Benefit for inversion of long-term satellite daily temperature based on multi-machine learning algorithms

The Advanced Very High-Resolution Radiometer (AVHRR), a satellite sensor has been in orbit for over 40 years, providing remote sensing images before 2000 and there is considerable room for improvement in the accuracy of current air temperature (T_a) inversions based on AVHRR to obtain accurate long-term T_a data before 2000. Here, we aim to estimate daily average (T_ave), maximum (T_max), and minimum (T_min) air temperatures at a resolution of 5 km for the Chinese region during the period 1983–2000. We developed a satellite-retrieval daily temperature extrapolation method based on machine learning (ML) combining multiple sources of big data, i.e., leveraging comprehensive and gap-free land surface temperature data from remote sensing along with other relevant variables from reanalysis data, topography and auxiliary data of local temperature, to generate extended time series of high-resolution T_a data. Quality validation results indicate that the ML can enhance the accuracy of T_a inversion with average error range of various ML methods being 0.995–1.606 °C, 1.316–1.971 °C and 1.396–1.980 °C for T_ave, T_max, and T_min, respectively, which is better than the 2.297 ± 1.704 °C, 3.294 ± 2.016 °C and 2.873 ± 1.666 °C of ERA5. Integrated ML method outperforms individual algorithms, yielding a high correlation coefficient of 0.96 and a robust mean error of 1 °C. The spatial distribution of newly daily T_a data from the multi-ML nationwide and local region is homogeneous with ERA5, indicating high physical consistency, and has higher resolution of 5 km. These updated temperature data can be beneficial in better revealing intricate structural attributes on a regional scale, as well as exploring urban heat islands.