Abstract:

Objective: In this study, an ensemble learning method based on Fengyun-4B geostationary meteorological satellite remote sensing data has been proposed to estimate spatially continuous near-surface air temperature (NSAT) in the Daxing’anling Mountains forest area, which can be used to provide data support and decision-making aids for forest fire risk assessment, early warning of drought in vegetation, and evaluation of ecological environment. Method: Taking the monitoring data from 20 meteorological stations around the Daxing’anling Mountains forest area as the true value of NSAT, combined with the remote sensing data from Chinese new FY-4B geostationary satellites, and based on the thermodynamic principle of influencing the NSAT, the land surface parameters （including the land surface temperature and land surface albedo）, the topographic parameters(including the slope, aspect , and the elevation）, as well as the spatiotemporal information（including the latitude, longitude, and the time of observation) were utilized to form the feature set, respectively. After preprocessing, such as spatiotemporal matching and normalization, the estimation model of NSAT in the Daxing’anling Mountains forest area is constructed and obtained using the gradient boosting decision tree ensemble learning algorithm. The feature set was randomly divided into the training set and the testing set in the ratio of 7:3, and the accuracy of the ensemble learning model was verified using the measured NSAT data from meteorological stations. The trained model was applied to estimate the near-surface air temperature in the Daxing’anling forest area, successfully obtaining the complete spatial distribution of near-surface air temperature in the region. Result: The results showed that the estimation model constructed in this study has an overall RMSE (root mean square error) of 1.393 ℃ for the training set and 1.621 ℃ for the test set, the model has no overfitting and underfitting phenomena, and 80% of the results error is less than 2.0 ℃, so it can accurately estimate the NSATs in the Daxing’anling Mountains forest area. It can be seen that in the NSAT estimation model, the feature parameter with the most significant influence on the weights is the land surface temperature. In addition, monitoring time, land surface albedo, latitude, and elevation are all essential feature parameters based on feature weight analysis. The prediction model was also applied to mapping the NSAT in the forest area, effectively obtaining the complete spatial distribution results. Conclusion: By combining the measured NSAT at meteorological stations with FY-4B geostationary satellite remote sensing image, it can effectively obtain NSAT images with short time intervals and continuous spatial distributions, which effectively makes up for the problem of data lacking that exists when only station measurements are utilized. For areas such as deep forests, which are rare and difficult to set up stations, it not only reduces the cost and difficulty of temperature monitoring but also further improves the efficiency and fineness, which has both theoretical and applied values.

Key words: near-surface air temperature, forest area, geostationary satellite, ensemble learning, Daxing’anling Mountains