Source-linked AI summary
Deep learning-based air temperature mapping by fusing remote sensing, station, simulation and socioeconomic data
Huanfeng Shen, Yun Jiang, Tongwen Li, Qing Cheng, Chao Zeng, Liangpei Zhang
TL;DR
Accurate Ta mapping is difficult because station coverage is sparse and Ta varies across complex spatial and temporal conditions. The study applies a five-layer DBN to fuse station, remote-sensing, assimilation, and socioeconomic data for mapping across China. The model achieves RMSE 1.996℃, MAE 1.539℃, and R 0.986 at the national scale.
Problem
Ta mapping is challenging because station distributions can be sparse, especially across large and complex regions, while Ta varies strongly across space and time.
Method
A five-layer DBN fuses station Ta observations with remotely sensed, assimilation, and socioeconomic data to model nonlinear relationships for high-resolution mapping.
Results
RMSE 1.996℃, MAE 1.539℃, and R 0.986 were achieved for national-scale mapping, with deep learning reporting the best overall performance against conventional methods.
Takeaways & Limitations
The results support the potential of deep learning for Ta estimation by incorporating nonlinear relationships and multiple data sources.
Takeaways & Limitations
Energy-balance approaches may require comprehensive input parameters that are difficult to obtain directly, while 16-day NDVI data can produce mapping outliers.
Abstract
from arXiv · showhide
Air temperature (Ta) is an essential climatological component that controls and influences various earth surface processes. In this study, we make the first attempt to employ deep learning for Ta mapping mainly based on space remote sensing and ground station observations. Considering that Ta varies greatly in space and time and is sensitive to many factors, assimilation data and socioeconomic data are also included for a multi-source data fusion based estimation. Specifically, a 5-layers structured deep belief network (DBN) is employed to better capture the complicated and non-linear relationships between Ta and different predictor variables. Layer-wise pre-training process for essential features extraction and fine-tuning process for weight parameters optimization ensure the robust prediction of Ta spatio-temporal distribution. The DBN model was implemented for 0.01° daily maximum Ta mapping across China. The ten-fold cross-validation results indicate that the DBN model achieves promising results with the RMSE of 1.996°C, MAE of 1.539°C, and R of 0.986 at the national scale. Compared with multiple linear regression (MLR), back-propagation neural network (BPNN) and random forest (RF) method, the DBN model reduces the MAE values by 1.340°C, 0.387°C and 0.222°C, respectively. Further analysis on spatial distribution and temporal tendency of prediction errors both validate the great potentials of DBN in Ta estimation.
1. Introduction
Air temperature mapping is difficult because Ta varies across space and time, station coverage is uneven, and LST is not a direct Ta proxy. The study introduces a five-layer DBN that fuses multi-source data for high-resolution Ta mapping across China.
- Ta estimation is challenging because sparse station distributions and complex environmental conditions limit spatial interpolation.
- LST cannot directly proxy Ta because the two variables have different physical meanings and magnitudes, while their relationship varies across space and time.
- Existing approaches include TVX, energy-balance, statistical, and machine-learning methods, each relying on distinct assumptions, physical mechanisms, or auxiliary variables.
- TVX methods can perform well in some settings but become infeasible in regions or seasons without high vegetation cover.
- Energy-balance methods offer physical interpretability and portability but may require comprehensive parameters that are difficult to obtain directly.
- The study’s five-layer DBN models nonlinear relationships between station Ta and remotely sensed, assimilation, and socioeconomic data for high-spatio-temporal-resolution mapping across China.
2. Study area and data
The study covers China’s highly heterogeneous terrain and climate using station observations together with remotely sensed, assimilation, socioeconomic, geographic, and temporal predictors. These data were harmonized, matched to a 0.01° grid, quality-controlled, and assembled into 107578 samples.
- 2.1. Study area: China spans approximately 9.6 million km2, with elevation ranging from -154 m to 8848 m and mountainous, hilly, and plateau areas covering about 67% of the land.
- 2.2. Data: The dataset includes 829 meteorological stations, with denser coverage in southern, eastern, and central China than in northwestern and southwestern regions.
- 2.2. Data: Daily maximum Ta observations from 2015 were combined with station geographic and temporal parameters.
- 2.2. Data: Predictors included MODIS LST, NDVI, land cover, elevation, GLDAS assimilation variables, road density, and population density.
- 2.2.5. Data pre-processing: All datasets were reprojected and resampled to 0.01°×0.01° grid cells, with raster values matched to station locations using nearest-neighbor extraction.
- 2.2.5. Data pre-processing: After removing unfilled and outlier records, 107578 matched samples containing Ta and all predictor values formed the experimental dataset.
3. Methodology
The study uses a deep belief network to model nonlinear relationships between air temperature observations and multi-source predictors. Its workflow combines unsupervised layer-wise pre-training, fine-tuning, cross-validation, and national-scale prediction.
- Deep belief network: A 5-layer deep belief network models nonlinear relationships between air temperature observations and multi-source data.The architecture includes one input layer, three hidden layers, and one output layer.
- Deep belief network: The DBN combines restricted Boltzmann machine layers with a back-propagation layer.Each RBM contains visible and hidden layers with fully undirected connections between them and no within-layer connections.
- Model training: Layer-wise unsupervised pre-training extracts essential features before back-propagation fine-tunes the DBN weights.Pre-training proceeds from the lowest to the highest hidden layer, after which estimated and observed Ta values guide refinement.
- Multi-source data fusion: Predictors include geographical and temporal parameters, remotely sensed data, assimilation data, socioeconomic data, and air temperature observations.The model represents Ta as a nonlinear function of variables including location, time, vegetation, soil moisture, albedo, population, and road density.
- Model evaluation: Ten-fold cross-validation with RMSE, MAE, and Pearson correlation evaluates predictive accuracy and overfitting.Each fold is withheld once for validation while the remaining nine groups fit the model.
4. Results
The results show that predictor relationships with Ta vary across environmental and temporal conditions, while DBN provides the strongest and most stable national-scale estimates among the compared models.
- 4.1. Descriptive statistics: LSTD has a strong correlation with Ta (R>0.9), while NDVI, Albe, Esoil, SoilMoi, Lat, Ele, PopD and RoadD show moderate correlations.
- 4.1. Descriptive statistics: Absolute R values below 0.2 indicate extremely weak correlations between Ta and WS, Vangle, Lon, Doy and Mon.
- 4.1. Descriptive statistics: LSTD-Ta correlation varies by circumstance, with nearly 0.4 R disparity between May and January and stronger relationships at low elevation, high latitude and cooler seasons.
- 4.2. Overall performance of the DBN model: DBN achieves the best national-scale cross-validation performance, with RMSE 1.996°C, MAE 1.539°C and R 0.986.
- 4.2. Overall performance of the DBN model: Compared with MLR, DBN decreases RMSE by 1.701°C and MAE by 1.340°C while increasing R by 0.037.
- 4.3. Spatial evaluation of model performance: Only 0.04% of stations have DBN estimation errors above 2°C, compared with 440, 37 and 23 stations exceeding 2.5°C for MLR, BPNN and RF.
- 4.4. Temporal evaluation of model performance: DBN residual medians remain close to 0°C across months, indicating less tendency toward overestimation or underestimation and lower residual uncertainty than other models.
5. Discussion
The discussion evaluates DBN structure, multi-source data fusion, uncertainty patterns, and spatial-temporal mapping behavior. It reports improved accuracy and detail, while identifying overfitting and misleading-variable effects that constrain reliability.
- DBN structure: Increasing DBN layers and neurons improved accuracy until four hidden layers, after which performance stabilized and over-parameterization risked overfitting.The selected architecture used three hidden layers with 25, 20, and 15 neurons, balancing performance and computational cost.
- Multi-source data fusion: Fusing multi-source datasets improved Ta estimation and provided more spatial detail in the mapping results.The combination R + S + A + P slightly reduced MAE compared with other tested combinations, while additional datasets optimized spatial information.
- Mapping uncertainties: The 16-day NDVI input produced mapping outliers because the model may not fit its regular temporal resolution to the daily scale.The discussion also identifies spatial overfitting as a risk for untrained or complicated locations.
- Spatio-temporal distributions: Seasonal Ta was higher in summer and lower in winter, with lower values over water and higher values in urban zones than surrounding rural areas.High-elevation regions such as the Tibetan Plateau were exceptions to the generally high summer temperatures.
6. Conclusions
The study applies a deep belief network to fuse multiple data sources for 0.01° daily maximum Ta mapping across China. It reports the best overall performance among compared methods, while noting unresolved challenges in variable selection, missing-data reconstruction, and regional or seasonal adaptation.
- Conclusions: A 5-layer DBN fused remote sensing, station, simulation and socioeconomic data to estimate spatially continuous 0.01° daily maximum Ta across China.The method was introduced as a deep learning approach for Ta estimation.
- Conclusions: RMSE 1.996°C, MAE 1.539°C and R 0.986 were achieved, representing the best overall performance among the compared methods.The validation results attribute the performance to better consideration of nonlinear relationships.
- Conclusions: DBN estimates provided more detailed spatio-temporal variations than assimilated Ta, and effective variables substantially improved model performance.The paper states that future work could extend the approach to daily minimum and mean Ta.
- Limitations: Scientifically selecting variables that may introduce mapping uncertainty remains challenging despite accuracy gains from multi-source datasets.The authors also identify incomplete LSTD coverage and data availability as areas requiring further exploration.
- Limitations: Future work should address missing information reconstruction and evaluate different models for specific regions and seasons.These directions are presented as ways to improve model performance.