Deep Learning-based Air Temperature Mapping By Fusing Multi-source Heterogeneous Data

In the context of increasing impact on the world derived from the global ecological environment problems,researches on climate change is becoming not only a subject matter,but also a social issue which attract wide concern.Air temperature(Ta),as an essential climatological component,plays a crucial role in the earth system,atmospheric system and physical process between the surface and atmosphere.The change of Ta controls and influences a broad range of environmental researches and applications,including human activities,climate change,crop growth,disease transmission and terrestrial hydrological cycle.Hence,accurate estimation of the spatial continuous distribution of Ta is extremely necessary for scientific cognition of the researches on ecological environment change and climate change.Traditionally,Ta is obtained by the meteorological station-based observations with the limited spatial coverage,which results in significant uncertainty in the regional scale researches.Recently,detailed knowledge of the spatial variability of Ta with high precision has been one of the challenges in related researches and applications.Meanwhile,satellite remote sensing provides the ability to extract valuable source of spatially continuous information,which can be used to estimate Ta and has developed a series of effective algorithms.However,there has not yet formed a universally applicable data and model for the Ta estimation in special areas.Comprehensive analysis also shows that there are still much room for improvement in the existing literature.For instance,existing models exhibit limited ability to mine the feature information of the data.Also,existing researches lack comprehensive considerations of the factors affecting the Ta.In addition,global model is generally used in previous literature,which ignores the spatio-temporal heterogeneity of Ta.In view of the above-mentioned issues,this paper takes the 0.01° daily maximum Ta across China as an example,and proposes a high precision Ta estimation method by combining multi-source data and deep learning theory.Besides,the spatio-temporal heterogeneity of Ta is taken into consideration,that is,establishing regional and seasonal models for the whole area to improve the accuracy.The specific research contents and conclusions are as follows:(1)Research on the deep learning-based Ta estimation by fusing multi-source heterogeneous data.This paper makes the first attempt to implement the deep belief network for Ta mapping and confirms the superiority of the overall and spatio-temporal model performance by comparing with traditional multiple linear regression,shallow neural networks and machine learning algorithms(sample-based overall model performance: RMSE=1.996°C,MAE=1.539°C,R=0.986).Then,different combinations of the multi-source datasets is evaluated to explore the substantial effect on the model accuracy.(2)Improving the above model by considering the regional and seasonal variety of Ta.Consider that Ta varies greatly in space and time,local models for the separated region and season are constructed based on the spatial and temporal attributes by using the clustering algorithm.Comparing with the experimental results of the global model,local models can significantly improve the estimation accuracy with the MAE value reducing by about 0.2? for specific region and season.(3)In-depth evaluations and analysis of Ta estimation results.This paper evaluates the overall and spatio-temporal model performance in detail by comparing with common methods.Besides,various of factors that affect the model accuracy are explored,such as the setting of model parameter,the combination of multi-source heterogeneous data and the distribution density of site-based Ta data.In addition,the model accuracy for the regional and seasonal variety considered Ta estimation model is compared detailedly with the original model.Then,Ta distribution with a spatial resolution of 0.01° in the country is mapped,and the uncertainties are also analyzed and solved.The mapping result exhibits more detailed spatial variations than the assimilated Ta data,especially for some complex area.The seasonal tendencies of mapping results are also obvious.