Research ArticleComparison of Machine-Learning Algorithms for Near-SurfaceAir-Temperature Estimation from FY-4A AGRI Data

Ke Zhou,1 Hailei Liu ,1,2 Xiaobo Deng,1,3 Hao Wang ,1 and Shenglan Zhang1

1Key Laboratory of Atmospheric Sounding, Chengdu University of Information Technology, Chengdu 610225, China2National Satellite Meteorology Center, China Meteorological Administration, Beijing 100081, China3Collaborative Innovation Center on Forecast and Evaluation of Meteorological Disasters,Nanjing University of Information Science & Technology, Nanjing 210044, China

Correspondence should be addressed to Hailei Liu; liuhailei@cuit.edu.cn

Received 21 April 2020; Revised 11 August 2020; Accepted 22 September 2020; Published 6 October 2020

Academic Editor: Stefania Bonafoni

Copyright © 2020Ke Zhou et al.'is is an open access article distributed under the Creative CommonsAttribution License, whichpermits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

Six machine-learning approaches, including multivariate linear regression (MLR), gradient boosting decision tree, k-nearestneighbors, random forest, extreme gradient boosting (XGB), and deep neural network (DNN), were compared for near-surfaceair-temperature (Tair) estimation from the new generation of Chinese geostationary meteorological satellite Fengyun-4A (FY-4A)observations. 'e brightness temperatures in split-window channels from the Advanced Geostationary Radiation Imager (AGRI)of FY-4A and numerical weather prediction data from the global forecast system were used as the predictor variables for Tairestimation. 'e performance of each model and the temporal and spatial distribution of the estimated Tair errors were analyzed.'e results showed that the XGB model had better overall performance, with R2 of 0.902, bias of −0.087°C, and root-mean-squareerror of 1.946°C.'e spatial variation characteristics of the Tair error of the XGBmethod were less obvious than those of the othermethods. 'e XGB model can provide more stable and high-precision Tair for a large-scale Tair estimation over China and canserve as a reference for Tair estimation based on machine-learning models.

1. Introduction

Air temperature (Tair) is one of the basic meteorological ob-servation parameters [1–3] and is of great concern in scientificdisciplines like hydrology, meteorology, and environmentalscience. Furthermore, it influences most land-surface pro-cesses, such as photosynthesis and land-surface evapotrans-piration [4]. Obtaining high-resolution Tair data can reducehuman health risks and promote urban heat island research, sohigh-resolution Tair information is quite crucial [5, 6]. 'esummer Tair value in China is generally above 20°C, except inthe high-altitude regions (e.g., Qinghai-Tibet Plateau). Sum-mer heat waves have a major impact on agricultural foodproduction, as well as the use of water and electricity [7]. 'isstudy focuses on the issue of summer Tair estimation in Chinausing Advanced Geostationary Radiation Imager (AGRI) data.

Large-scale Tair data are mainly obtained by interpolationfrom the data collected by surface meteorological stations.

However, the distribution of meteorological stations is usuallyuneven due to geographical factors, and some sparsely pop-ulated areas even have no meteorological observation [8].'erefore, the accuracy of the interpolated Tair data is limited,and researchers are unable to obtain high-spatial-resolutionTair information [9].

Meteorological satellites such as low-Earth-orbit (LEO)satellites and geostationary-Earth-orbit (GEO) satellites canprovide continuous surface (i.e., land-surface temperature(LST)) and atmospheric observations with a wide spatialcoverage at global and regional scales [10–12]. In the lastseveral decades, LEO and GEO observations have beengradually applied to Tair estimation with the development ofmeteorological satellite technology. LEO satellites can onlyacquire data once or twice a day for one place. In addition,cloud contamination will reduce the effective data for Tairestimation [13–15]. Unlike LEO satellites, GEOmeteorologicalsatellites can continuously provide data every 15 or 30min on

HindawiAdvances in MeteorologyVolume 2020, Article ID 8887364, 14 pageshttps://doi.org/10.1155/2020/8887364

mailto:liuhailei@cuit.edu.cnhttps://orcid.org/0000-0003-4693-061Xhttps://orcid.org/0000-0003-0090-2840https://creativecommons.org/licenses/by/4.0/https://doi.org/10.1155/2020/8887364

one-third of the Earth’s surface [16–20]. 'erefore, GEOsatellites comprise an effective method of obtaining high-spatial- and high-temporal-resolution Tair data in a fixed areaand have the potential to facilitate the study on the dailychange of Tair [20, 21].

At present, the methods for Tair estimation from satellitebrightness temperatures (BTs) and land-surface temperature(LST) product data can be divided into simple linear,multivariate linear, and nonlinear approaches [21, 22].Previous studies [7, 23, 24] have shown that machine-learning algorithms can obtain higher-accuracy Tair valuesthan those in other methods. For example, a machine-learning model (e.g., a neural network model (NN)) hashigher accuracy, and the root-mean-square error (RMSE) isreduced by 1.29°C compared with linear models [7].

'e AGRI aboard Fengyun-4A (FY-4A) has 14 spectralbands [18, 20, 25, 26]—six visible/near-infrared (VIS/NIR),six infrared (IR), and two water vapor bands—with atemporal resolution of 15min for the full disk and a spatialresolution of 4 km at IR bands. It provides an unprecedentedopportunity for obtaining high-precision Tair data overChina and surrounding areas.

Machine-learning methods are used to estimate Tairbased on moderate-resolution imaging spectroradiometer(MODIS) data in several studies [27–29]. However, there iscurrently a lack of relevant studies on Tair estimation basedon FY-4A.'e use of FY-4A data to estimate high-resolutionTair is of great significance to the study of human health andhigh-temporal- and high-spatial-resolution Tair in East Asia.In addition, there is a need for timely and high-resolutionTair data for the sustainable planning and management ofclimate-resilient cities [3].

'is study aims to develop the machine-learning ap-proaches for Tair estimation using FY-4A data and comparesthe performances of different machine-learning models [i.e.,multivariate linear regression (MLR), gradient boostingdecision tree (GBTD), k-nearest neighbors (KNN), randomforest (RF), extreme gradient boosting (XGB), and deepneural network (DNN)] in Tair estimation, which, to the bestof our knowledge, has never been done before. By comparingdifferent machine-learning algorithms, a machine-learningalgorithm with good applicability for estimating Tair is se-lected. 'e algorithm is widely applicable to meteorologicalsatellites without surface-temperature products.

'e remainder of this paper is organized as follows. InSection 2, the study area and data used for model devel-opment are introduced, and the construction of the above-listed six machine-learning models for Tair estimation isdescribed. Variable importance analysis, validation results,and discussion are described in Section 3. Conclusions arepresented in Section 4.

2. Materials and Methods

2.1. Study Area. 'e study area is located in China, andFigure 1 shows the spatial distribution of 1,812 meteoro-logical stations used in this study. 'ere is a higher altitudein the West over China than in the East, and even theQinghai-Tibet Plateau has an average elevation of over

4,000m [30]. 'ere are more stations in the East areas thanin the West ones due to the uneven distribution of pop-ulation and economic development in China (Figure 1).

2.2.Data. 'edata used in this studymainly include FY-4A/AGRI brightness temperature (BT) and L2 cloud mask data,global forecast system (GFS) 3 h forecast data, meteoro-logical data of 1,812 stations in China, and other auxiliarydata (longitude, latitude, and Julian day).

2.2.1. Satellite Data. FY-4A, the new generation of Chinesegeostationary meteorological satellites, was launched onDecember 11, 2016. It was fixed at a position of 99.5°E abovethe equator. As thermal infrared split-window channels, the12 and 13 bands of AGRI (BT12 and BT13, respectively) aremainly used for studies of cloud, aerosol, and Tair estimation.'eir central wavelengths are 10.8 and 12.0 μm [31].

BT12, BT13, and L2 cloud mask products during Summer2018 (i.e., June, July, and August) were used. 'e ARGI datawere selected at 3 h intervals (i.e., 00, 03, 06, 09, 12, 15, 18,and 21 UTC) per day. 'e data were downloaded from theChina National Satellite Meteorological Center (http://satellite.nsmc.org.cn/PortalSite/Data/Satellite.aspx).

2.2.2. Meteorological Data. 'is study selected meteoro-logical data at 3 h intervals from 1,812 observation stations inChina during summer 2018. 'e meteorological variablesused in this study include Tair and the digital elevation model(DEM). Tair in summer 2018 ranges from −5°C to 40°C, andthe DEM of the station was between 0 and 5000m. 'esedata were obtained from the China Meteorological DataService Center (CMDC) (http://data.cma.cn/).

2.2.3. Numerical Weather Prediction Data and AuxiliaryData. Previous studies showed that the relationship be-tween BTs (or LST) and Tair is easily affected by surfacecharacteristics and atmospheric conditions [7, 31]. 'ere-fore, the accuracy of Tair estimation was effectively improvedby adding several auxiliary parameters [32]. In this study,GFS 3 h precipitable water vapor (GFS PWV) and relativehumidity (GFS RH) forecast fields data were used. 'eforecast length of the GFS data (GFS PWV and GFS RH)used was 3 h per day, and there were eight periods of data perday (i.e., 00, 03, 06, 09, 12, 15, 18, and 21 UTC).'e GFS datawere interpolated according to the location and time in-formation of the AGRI pixels. GFS data were obtainedthrough the U.S. National Oceanic and Atmospheric Ad-ministration (NOAA) National Centers for EnvironmentalPrediction (http://www.nco.ncep.noaa.gov/pmb/products/gfs). Table 1 presents the temporal and spatial resolutioninformation of the data used in this study.

2.3. Methods

2.3.1. Preparation of Training Dataset. 'e BT12, BT13, GFSPWV, GFS RH, and auxiliary data were used as the inputvariables, Tair was used as the response variable of the

2 Advances in Meteorology

http://satellite.nsmc.org.cn/PortalSite/Data/Satellite.aspxhttp://satellite.nsmc.org.cn/PortalSite/Data/Satellite.aspxhttp://data.cma.cn/http://www.nco.ncep.noaa.gov/pmb/products/gfshttp://www.nco.ncep.noaa.gov/pmb/products/gfs

machine-learning models (Table 1), and all data points(across space and time) were included in one model (i.e., theXGB model) [33]. 'e construction of the representativetraining data was crucial to develop successful retrievalmodels using machine learning. 'us, data from June toAugust—except the 1st, 10th, 20th, and 30th of eachmonth—were collected as the original dataset, and theoriginal dataset was randomly divided into a training dataset(80%, 97,1773 samples) and a test dataset (20%, 24,2944samples) with the same number of pieces of data for each bin(i.e., 1.0°C in temperature) as shown in Figure 2. For thevalidation, the data that were not used for training wereselected from June to August 1st, 10th, 20th, and 30th.

2.3.2. Machine-Learning Algorithm. Machine-learningmethods have been widely used in classification and re-gression in the field of remote sensing [34–41]. In this study,six machine-learning approaches, that is, MLR, GBTD,KNN, RF, XGB, and DNN, were used for constructing Tairestimation models. 'e flowchart of Tair estimation based onmachine-learning approaches is shown in Figure 3. L2 cloud

mask products were used to detect cloud. If the data werecloudless, FY-4A data matched both the GFS data andmeteorological station data (same space and time), and thenTair was estimated through the machine-learning models.

As a simple machine-learning algorithm, MLR hasusually been the basic tool for the estimation of meteo-rological parameters [42, 43]. Similarly, as a local non-linear algorithm, the prediction process of KNN isgenerally divided into two steps. First, when the KNNalgorithm predicts a point, it searches for the k-nearestneighbors closest to the point in the training dataset.Second, the mean of the target variable of the k-nearestneighbors is computed [44, 45]. In this study,the hyperparameters of MLR and KNN were set to defaultvalues. Unlike MLR and KNN, RF is an ensemble to adecision-tree-based approach for improving the predic-tion accuracy, such that each tree depends on the values ofa random vector sampled independently and withthe same distribution for all trees in the forest[34, 43, 46–50]. 'e Scikit-learn library was used forhyperparameter tuning named GridSearchCV from Py-thon to filter the hyperparameters including number of

80°0′0″E

N

0 140 280 560 Miles

20°0′0″

N30

°0′0″

N40

°0′0″

N50

°0′0″

N

90°0′0″E 100°0′0″E 110°0′0″E 120°0′0″E 130°0′0″E

4000 - 50003000 - 39992000 - 2999

1000 - 19990 - 999

Figure 1: Elevation of study area and meteorological sites in China. Differently colored dots represent sites located at different elevations.

Table 1: Temporal and spatial information of the primary data used in this study.

Abbreviation Units Spatial resolution Temporal resolution SourceAGRI BTs K 4 km 15min FY-4A AGRIGFS PWV mm 0.5° 3°h GFSGFS RH % 0.5° 3°h GFSDEM m Site — FY-4A AGRILongitude — Site — CMDC/AGRILatitude — Site — CMDC/AGRITair °C Site 1°h CMDCJulian day — — — —

Advances in Meteorology 3

All datasets

0

20000

40000

60000

80000

100000

Num

ber o

f dat

a

0 5 10 15 20 25 30 35 40 45–5Temperature (°C)

(a)

Selected training dataset

0

20000

30000

40000

50000

60000

70000

80000

Num

ber o

f dat

a

0 5 10 15 20 25 30 35 40 45–5Temperature (°C)

(b)

Selected test dataset

0

2500

5000

7500

10000

12500

15000

17500

Num

ber o

f dat

a

0 5 10 15 20 25 30 35 40 45–5Temperature (°C)

(c)

Figure 2: Histograms of training data for machine learning. (a) All datasets, (b) training dataset, and (c) test dataset.

Machine learning models(MLR, RF, GBTD, KNN, XGB, DNN)

FY-4A/AGRI data (BT12, BT13)

Clear sky?(L2 Product)End

No

Yes

Machine learning models(MLR, RF, GBTD, KNN, XGB, DNN)

Estimation ofairtemperature

CMDC data (Tair, Julian day,longitude, latitude)

GFS data (GFS PWV, GFSRH)

Figure 3: Flowchart of the Tair estimation model in this study.

4 Advances in Meteorology

trees (n_estimators), minimum number of samples(min_samples_leaf ), and maximum depth of a tree(max_depth). 'e result of parameter selection isn_estimators � 200, min_samples_leaf � 50, andmax_depth � 3.

'e principle of GBTD is to sequentially apply a clas-sification algorithm to the weighted version of the trainingdata [51, 52], descending along the gradient direction of themodel loss function previously established, and then per-form a weighted majority vote on the resulting classifiersequence. As an improved algorithm of GBTD, XGB uses alldata in each iteration, which is similar to RF [53, 54].'erefore, XGB reduces the complexity of the model andmakes the learned model simpler [35, 54–58]. In this study,four hyperparameters in GBTD and XGB models (i.e.,n_estimators, max_depth, learning_rate (lr), and minimumloss reduction) required to make a further partition on a leafnode of the tree (gamma) were empirically tuned based onRMSE. 'e optimum n_estimators, gamma, max_depth,and lr in the two models were 500, 0.2, 5, and 0.1,respectively.

An artificial neural network (ANN) is a biologicallyinspired machine-learning method [59]. Here, DNN, asubset of ANN with multiple hidden layers, uses a fullyconnected structure, which has the ability to learn time andspace relationships [60, 61]. It adjusts the connectionstrength through back-propagation and minimizes theprediction error by iterating between neurons [62–64]. Eachhidden layer was tested in the DNN model at one to fivehidden layers and 5–200 neurons in five intervals. In ad-dition, some widely used optimizers (i.e., stochastic gradientdescent, RMSProp, and Adam) were tested by comparing thecalculated results. In this study, the hyperparameters of theDNN were set as follows: batch_size, 128; dropout_rate, 0.1;stop_steps, 20 (if the validation-set loss function was notimproved within 20, training will be terminated); andlearning rate, 0.001. 'e optimizer chose Adam, the numberof hidden layers was three, and the number of hiddenneurons was 256.

2.4. Error Analyses. Four statistical factors—determinationcoefficient (R2), RMSE, MSE, and mean bias (bias)—wereused to evaluate the accuracy of Tair estimation model asfollows:

R2

� Tea − Tea( Toa − Toa(

2

Tea − Tea( 2

Toa − Toa( 2 , (1)

RMSE �

�������������

N

i�1Tea − Toa(

2

N

,(2)

bias �

N

i�1Tea − Toa(

N,

(3)

MSE �

N

i�1Tea − Toa(

2

N,

(4)

where Tea is the estimated Tair, Toa is the observed Tair at themeteorological stations, and N is the sample size.

3. Results and Discussion

In this section, the results of variable importance werepresented, and the performance of the six machine-learningmodels was verified. 'e spatial distribution characteristicsof the Tair errors of each model were also analyzed.

3.1. Variable Importance Results. Correlation analysis wasperformed to analyze the linear relationship between Tair andBT12, BT13, GFS PWV, GFS RH, DEM, longitude (LONG),latitude (LAT), and Julian day (JD). Table 2 shows thecorrelation coefficient matrix of these variables.

As described in Figure 4(a), GFS PWV, DEM, BT12, andBT13 had a better correlation with Tair than other variables,and the R values of the four variables were 0.635, −0.596,0.459, and 0.413, respectively. 'is indicated that thesevariables played more important roles in the linear Tairestimation models. However, the Pearson correlation co-efficient only described the linear correlation between twovariables; it could not identify the nonlinear relationshipbetween two variables. 'erefore, the variable importance ofthe RF algorithm was also analyzed (Figure 4(b)). 'e RFalgorithm modeled the nonlinear relationship well. 'e GFSPWV was identified as the most important variable for Tairestimation in the RF model, while the GFS RH and BT12 alsoplayed important roles than other predictors. 'erefore,PWV and RH were used as inputs to effectively improve theaccuracy of Tair estimation, which was consistent with theprevious study [65].

3.2. Model Performance Results. For evaluating the overallperformance of each model, a 10-fold cross-validationmethod was used. K-fold cross-validation was used formodel configuration selection. When a particular value of Kwas selected (where K was 10), the datasets were randomlyand equally distributed among K groups. One group wasfolded for test, and the K− 1 group was folded for training.In a total of k validations, the model performance wascalculated using different test folds for each validation [35].Finally, the average validation results were used to evaluatethe overall performance of each model.

Figure 5 illustrates the six models with different sta-tistical parameters, including RMSE, Bias, MSE, and R2.'e MLR model had the lowest performance of the sixmodels. 'e variation range of RMSE, Bias, MSE, and R2 inthe MLR model was quite wide; even the range of RMSEwas 1.602°C–4.487°C, while the DNN model used in thisstudy had better overall performance and higher efficiencythan the other five models. 'e DNN model showed thehighest accuracy, with an average RMSE of 1.736°C. 'e

Advances in Meteorology 5

range of RMSE in the DNN model was 0.852°C–2.584°C,showing good concentration and stability, as presented inFigure 5(a). In addition, the overall performances of theXGB and GBTD models of the remaining models wereequivalent, which were better than those of the MLR, KNN,and RF models.

3.3. Validation Results. Model performance was used as anindicator to internally validate each model. 'e model ac-curacy must be evaluated with a dataset that was not used fortraining or testing. To validate the developed MLR, RF,KNN, GBTD, XGB, and DNNmodels, the observed data notused for both training and testing were utilized (validationdataset in Section 2.3.1). Figure 6 illustrates the quantitativevalidation results of the estimated Tair during the validationtime (the 1st, 10th, 20th, and 30th of June–August 2018).Compared with the results in the test dataset, the overallaccuracy of the six models on the validation dataset de-creased. For example, for the DNN model, the RMSE of Tairusing the test dataset was 1.736°C, while that of the vali-dation results was 2.006°C. 'is difference may be caused byoverfitting due to the fact that the best model was not se-lected based on the final validation results [35].

'e biases of the MLR, RF, DNN, GBTD, and XGBmodels were within ±0.2°C, indicating no obvious overes-timation or underestimation. In contrast, the KNN modelshowed a larger negative bias of −0.492°C. 'e reason thatthe KNNmodel had a larger negative bias may be that it hadpoor robustness. Robustness mainly depended on thedataset, and poor robustness made the model difficult todirectly apply to other cases, so the KNN model had a lowbias on the test dataset and a high bias on the validationdataset.

'e XGB model had excellent modeling performancewith R2 of 0.902. 'e R2 values of the GBTD and DNNmodels were 0.898 and 0.890, respectively, and the R2 valueof the remaining three models was less than 0.89. Moreover,compared with the other models, the XGB and GBTDmodels can repeatedly learn to generate a weighted averageof the weak learners. 'erefore, the XGB and GBTD modelsshowed a relatively better performance in the validationdataset in most sites. In general, the XGB model showed ahigher overall performance than the other five models on thevalidation dataset.

'e Tair estimation models based on satellite and nu-merical forecast data are susceptible to factors such as al-titude and surface roughness. To further evaluate the

Table 2: Pearson correlation matrix for variables considered in the Tair estimation model.

Tair BT12 BT13 GFS PWV GFS RH DEM LONG LAT JDTair 1.000 0.459 0.413 0.635 −0.182 −0.596 0.303 −0.288 −0.383BT12 0.459 1.000 0.995 0.047 −0.256 −0.287 0.199 −0.022 −0.099BT13 0.413 0.995 1.000 0.010 −0.249 −0.268 0.190 0.002 −0.077GFS PWV 0.635 0.047 0.010 1.000 0.383 −0.585 0.366 −0.463 −0.310GFS RH −0.182 −0.256 −0.249 0.383 1.000 −0.020 0.189 −0.355 −0.046DEM −0.596 −0.287 −0.268 −0.585 −0.020 1.000 −0.663 0.057 0.003LONG 0.303 0.199 0.190 0.366 0.189 −0.663 1.000 0.134 −0.004LAT −0.288 −0.022 0.002 −0.463 −0.355 0.057 0.134 1.000 −0.006JD −0.383 −0.099 −0.077 −0.310 −0.046 0.003 −0.004 −0.006 1.000Data represent the correlation coefficient between different variables.

0.413

0.459

–0.182

0.63.5

–0.596

–0.288

0.303

–0.383JD

LON

LAT

DEM

PWV

RH

BT12

BT13

–0.8 –0.6 –0.4 –0.2 0 0.2 0.4 0.6 0.8 1.0–1.0Correlation coefficient

(a)

0.012

0.165

0.165

0.444

0.080

0.048

0.021

0.065JD

LON

LAT

DEM

PWV

RH

BT12

BT13

0.2 0.4 0.6 0.8 1.00Attribute usage

(b)

Figure 4: (a) Relative variable importance identified by Pearson correlation coefficient and (b) RF variable importance.

6 Advances in Meteorology

applicability of these models, the spatial distribution of eachmeteorological observation was evaluated (Figures 7–9).

It can be seen that the Tair estimation errors of all modelsshowed obvious spatial distribution characteristics (Fig-ure 7). Generally, the RMSE is relatively low in the easternregions (e.g., Guangdong Province) and high in thenorthwestern regions for each model (e.g., Xinjiang Prov-ince). For example, the RMSE in Guangdong Province of theXGB model was approximately 1.2°C–1.8°C, while that inXinjiang Province was about 2.0°C–3.2°C. Because thenorthwestern regions have relatively wide Tair changesduring day and night, high altitude, and few meteorologicalobservations, the accuracy difference between northwesternand eastern China is obvious. Moreover, the RMSE of theKNN, DNN, GBTD, and XGB models was relatively low inthe eastern and southern regions. However, the MLR, RF,KNN, and DNNmodels had a higher RMSE in northwesternChina. In contrast, the GBTD and XGB models had a rel-atively smaller RMSE in northwestern China because the

GBTD and XGB models can generate repeated weightedaverages to adjust the applicability of different regionsthrough repeated learning of numerous data.

Furthermore, Gong’s study (2015) [66] illustrated thatthe RMSE of GFS Tair in most eastern regions reaches1.5°C–3.0°C and was above 3.5°C in the northwestern re-gions. By contrast, the results showed that the RMSE of Tairestimated by the DNN, XGB, and GBTD models was ob-viously lower than that of GFS data. In the present study, theRMSE of the XGB model was 1.0°C–2.0°C in most easternregions, and it was below 3.5°C in the northwestern regions.In addition, RMSE< 2.0°C accounted for 48.2% andRMSE< 2.5°C accounted for 87.6% in the XGB model.

'e six models showed the same distribution trend asshown in Figure 8, with R2 being higher in the easternregions, but R2 gradually became lower as it got closer to thesouthwestern regions. Compared with the central regions(e.g., Henan Province), the viewing zenith angle (VZA) ofARGI over the western China is larger. 'e larger the VZA

0

1

2

3

4

5RM

SE (°

C)

KNN GBTD RF XGB DNNMLRModel

(a)

–3

–2

–1

0

1

2

3

Bias

(°C)

KNN GBTD RF XGB DNNMLRModel

(b)

0

2

4

6

8

10

12

14

16

18

MSE

KNN GBTD RF XGB DNNMLRModel

(c)

0.0

0.2

0.4

0.6

0.8

1.0

1.2

R2

KNN GBTD RF XGB DNNMLRModel

(d)

Figure 5: Boxplots of performance evaluation for the six models (MLR, KNN, SVM, RF, XGB, and DNN) using the test dataset in terms of(a) RMSE, (b) Bias, (c) MSE, and (d) R2.

Advances in Meteorology 7

4035302520

Estim

ated

Tai

r (°C

)

1510

50

–5–5 0 5 10 15

Station Tair (°C)20 25 30 35 40

400

350

300

250

200

150

100

50

0

RMSE = 2.840°CBias = 0.055°CR2 = 0.791

(a)

Station Tair (°C)–5 0 5 10 15 20 25 30 35 40

4035302520

Estim

ated

Tai

r (°C

)

1510

50

–5

400

350

300

250

200

150

100

50

0

RMSE = 2.233°CBias = –0.185°CR2 = 0.878

(b)

Station Tair (°C)–5 0 5 10 15 20 25 30 35 40

4035302520

Estim

ated

Tai

r (°C

)

1510

50

–5

400

350

300

250

200

150

100

50

0

RMSE = 2.120°CBias = –0.492°CR2 = 0.884

(c)

Estim

ated

Tai

r (°C

)

40353025201510

50

–5

Station Tair (°C)–5 0 5 10 15 20 25 30 35 40

400

350

300

250

200

150

100

50

0

RMSE = 2.006°CBias = 0.185°CR2 = 0.890

(d)

Station Tair (°C)–5 0 5 10 15 20 25 30 35 40

4035302520

Estim

ated

Tai

r (°C

)

1510

50

–5

400

350

300

250

200

150

100

50

0

RMSE = 1.984°CBias = –0.122°CR2 = 0.898

(e)

Station Tair (°C)–5 0 5 10 15 20 25 30 35 40

4035302520

Estim

ated

Tai

r (°C

)

1510

50

–5

400

350

300

250

200

150

100

50

0

RMSE = 1.946°CBias = –0.087°CR2 = 0.902

(f )

Figure 6: Two-dimensional histogram of predicted Tair data and meteorological observed Tair data based on six machine-learning models.(a) MLR model. (b) RF model. (c) KNN model. (d) DNN model. (e) GBTD model. (f ) XGB model.

50°N

40°N

30°N

20°N

80°E 90°E 100°E 110°ELatitude

120°E 130°E

Long

itude

0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.00.0Temperature (°C)

(a)

Latitude80°E 90°E 100°E 110°E 120°E 130°E

50°N

40°N

30°N

20°N

Long

itude

0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.00.0Temperature (°C)

(b)

Latitude80°E 90°E 100°E 110°E 120°E 130°E

50°N

40°N

30°N

20°N

Long

itude

0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.00.0Temperature (°C)

(c)

50°N

40°N

30°N

20°N

80°E 90°E 100°E 110°ELatitude

120°E 130°E

Long

itude

0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.00.0Temperature (°C)

(d)

Latitude80°E 90°E 100°E 110°E 120°E 130°E

50°N

40°N

30°N

20°N

Long

itude

0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.00.0Temperature (°C)

(e)

Latitude80°E 90°E 100°E 110°E 120°E 130°E

50°N

40°N

30°N

20°N

Long

itude

0.5 1.0 1.5 2.0 2.5 3.0 3.5 4.00.0Temperature (°C)

(f )

Figure 7: Spatial distribution of RMSE for six machine-learning models. (a) MLR-RMSE. (b) RF-RMSE. (c) KNN-RMSE. (d) DNN-RMSE.(e) GBTD-RMSE. (f ) XGB-RMSE.

8 Advances in Meteorology

is, the more the radiation reaching the sensor will be highlyaffected by the atmosphere, which may cause differences inR2 of the estimated Tair value between the southwestern andcentral regions.

For the MLR model, the bias for all of China was large.For the RF and KNN models, relatively high negative biasexisted in southwestern China (e.g., Yunnan-Guizhou Pla-teau), as shown in Figure 9. 'is may be the relatively simplestructure of the three models mentioned above, whichcannot well simulate the complex Tair changes in China,resulting in underfitting. Besides, Tair estimated by the DNNmodel was overestimated in northwestern China, which wasthe reason that the RMSE in the DNNmodel was also high inthese regions. In contrast, the GBTD and XGB models hadrelatively low bias in northwestern China, where the absolutebias ranges from 2.0°C to 3.0°C. In conclusion, the bias islower in the coastal areas and higher in northwestern areas,which is mainly related to the characteristics of Summer Tairchange.

Figure 10 shows the time series of RMSE for the sixmodels during the validation period. 'e RMSE of theMLR model was significantly higher than other models,with the RMSE ranging from 2.5°C to 4.3°C. In contrast,the RMSE of the GBTD and XGB models showed arelatively lower RMSE (i.e., 1.8°C–2.2°C) than that in theRF, KNN, and DNN models.

Based on the above analysis, it is expected that theXGB model can provide a more reliable and accurate Tairestimation than other models. For purposes of evaluatingthe contribution of predictive factors in the XGB model

to Tair estimation, BTs data (BT12 and BT13) and GFS data(GFS PWV and RH) were successively introduced (Ta-ble 3). As shown in Table 3, DEM, longitude, latitude, andJulian day were used as input variables, and the RMSE ofthe XGB model was 3.003°C. 'e accuracy of Tair esti-mation was obviously improved when BT12 and BT13 wereincluded in the model. Moreover, when GFS PWV andRH were added to the input variables, the RMSE of theXGB model was decreased to 2.164° C, indicating im-portant influences of GFS PWV and RH on the Tair es-timation. 'ese results are understandable due to the factthat PWV and RH are the main parameters needed foratmospheric correction and LST retrieval. 'e RMSE ofXGB model was improved by 0.228° C compared with justGFS data which were introduced when both AGRI BTsand GFS data were introduced to the input variables. 'isindicates that both GFS data and satellite observationdata have an important role in improving the Tair esti-mation model. 'e RMSE of Tair estimation model wasless than 2.0° C when both satellite BTs and GFS data wereintroduced, which was considered to be the precisionlevel of “accurate” [67].

'e relationship of XGB model errors with altitude,observed Tair, and VZA was analyzed. Figure 11 dem-onstrates the scatter plot of the estimated Tair error withDEM, Tair, and VZA. It can be seen that the Tair errormainly ranges from −3°C to 3°C. 'e results showedpositive deviation at high-altitude areas, which produceda larger RMSE than low-altitude areas. 'e model showeda positive deviation when Tair was low while exhibiting a

50°N

40°N

30°N

20°N

80°E 90°E 100°E 110°ELatitude

120°E 130°E

Long

itude

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.00.0

(a)

Latitude80°E 90°E 100°E 110°E 120°E 130°E

50°N

40°N

30°N

20°N

Long

itude

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.00.0

(b)

Latitude80°E 90°E 100°E 110°E 120°E 130°E

50°N

40°N

30°N

20°N

Long

itude

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.00.0

(c)

80°E 90°E 100°E 110°ELatitude

120°E 130°E

50°N

40°N

30°N

20°N

Long

itude

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.00.0

(d)

Latitude80°E 90°E 100°E 110°E 120°E 130°E

50°N

40°N

30°N

20°N

Long

itude

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.00.0

(e)

Latitude80°E 90°E 100°E 110°E 120°E 130°E

50°N

40°N

30°N

20°N

Long

itude

0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.00.0

(f )

Figure 8: Spatial distribution of R2 for six machine-learning models. (a) MLR-R2. (b) RF-R2. (c) KNN-R2. (d) DNN-R2. (e) GBTD-R2.(f ) XGB-R2.

Advances in Meteorology 9

80°E 90°E 100°E 110°ELatitude

120°E 130°E

50°N

40°N

30°N

20°N

Long

itude

–2.5 –2.0 –1.5 –1.0 –0.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0–3.0Temperature (°C)

(a)

Latitude80°E 90°E 100°E 110°E 120°E 130°E

50°N

40°N

30°N

20°N

Long

itude

–2.5 –2.0 –1.5 –1.0 –0.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0–3.0Temperature (°C)

(b)

Latitude80°E 90°E 100°E 110°E 120°E 130°E

50°N

40°N

30°N

20°N

Long

itude

–2.5 –2.0 –1.5 –1.0 –0.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0–3.0Temperature (°C)

(c)

80°E 90°E 100°E 110°ELatitude

120°E 130°E

50°N

40°N

30°N

20°N

Long

itude

–2.5 –2.0 –1.5 –1.0 –0.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0–3.0Temperature (°C)

(d)

Latitude80°E 90°E 100°E 110°E 120°E 130°E

50°N

40°N

30°N

20°N

Long

itude

–2.5 –2.0 –1.5 –1.0 –0.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0–3.0Temperature (°C)

(e)

Latitude80°E 90°E 100°E 110°E 120°E 130°E

50°N

40°N

30°N

20°N

Long

itude

–2.5 –2.0 –1.5 –1.0 –0.5 0.0 0.5 1.0 1.5 2.0 2.5 3.0–3.0Temperature (°C)

(f )

Figure 9: Spatial distribution of bias for six machine-learning models. (a) MLR-Bias. (b) RF-Bias. (c) KNN-Bias. (d) DNN-Bias. (e) GBTD-Bias. (f ) XGB-Bias.

2018

/06/

01

2018

/06/

10

2018

/06/

20

2018

/06/

30

2018

/07/

10

Date

RMSE between estimated Tair and observed Tair

2018

/07/

20

2018

/07/

30

2018

/08/

10

2018

/08/

20

2018

/08/

30

MLRRFKNN

DNNGBTDXGB

1.5

2.0

2.5

3.0

3.5

4.0

4.5

RMSE

(°C)

Figure 10: Time series (June to August 1st, 10th, 20th, and 30th) of RMSE of estimated Tair for six machine-learning models.

Table 3: 'e contribution of AGRI BTs and GFS data to the XGB Tair estimation model.

Predictive factors XGB modelRMSE (°C)DEM, longitude, latitude, Julian day 3.003°CDEM, longitude, latitude, Julian day, BT12, BT13 2.376°CDEM, longitude, latitude, Julian day, GFS PWV, GFS RH 2.164°CDEM, longitude, latitude, Julian day, BT12, BT13, GFS PWV, GFS RH 1.946°C

10 Advances in Meteorology

negative bias for the high-air-temperature condition.'erefore, the model showed a larger RMSE in the lower-and higher-air-temperature conditions due to underes-timation and overestimation. 'is is similar to the resultsof previous studies [38]. Furthermore, the uneven dis-tribution of stations makes the applicability of the modelin high-altitude areas poor. It is worth mentioning thatthe effect of VZA on model performance is negligible asshown in Figure 11(c).

4. Conclusions

In this study, six machine-learning approaches (MLR, RF,KNN, DNN, GBTD, and XGB) for Tair estimation from FY-4A AGRI data in China were compared and analyzed interms of the spatial and temporal characteristics of theirperformance. 'e validation results highlighted the highpotential of Tair estimation approaches using machinelearning and showed that the accuracy of the XGB modelwas better than that of theMLR, RF, KNN, GBTD, and DNNmodels at most sites for Tair estimation over China. 'evalidation was performed using spatially and temporallyindependent data, and hence the model performance wasconsidered to be quite reliable.

'is study improves on previous studies in the followingkey areas. First,Tair estimationmodels were constructed basedon FY-4A AGRI data and other auxiliary data. 'e resultsshowed that high-temporal- and high-spatial-resolution Tairvalues (RMSE

[2] L. Prihodko and S. N. Goward, “Estimation of air temperaturefrom remotely sensed surface observations,” Remote Sensingof Environment, vol. 60, no. 3, pp. 335–346, 1997.

[3] Z. S. Venter, O. Brousse, I. Esau, and F. Meier, “Hyperlocalmapping of urban air temperature using remote sensing andcrowdsourced weather data,” Remote Sensing of Environment,vol. 242, no. 1, Article ID 111791, 2020.

[4] H. C. Ho, A. Knudby, Y. Xu, M. Hodul, and M. Aminipouri,“A comparison of urban heat islands mapped using skintemperature, air temperature, and apparent temperature(Humidex), for the greater Vancouver area,” Science of theTotal Environment, vol. 544, pp. 929–938, 2016.

[5] T. R. Lookingbill and D. L. Urban, “Spatial estimation of airtemperature differences for landscape-scale studies in mon-tane environments,” Agricultural & Forest Meteorology,vol. 114, no. 3-4, pp. 141–151, 2003.

[6] J. W. Hurrell and K. E. Trenberth, “Satellite versus surfaceestimates of air temperature since 1979,” Journal of Climate,vol. 9, no. 9, pp. 2222–2232, 1996.

[7] H. Liu, Q. Zhou, S. Zhang, and X. Deng, “Estimation ofsummer air temperature over China using himawari-8 AHIand numerical weather prediction data,” Advances in Mete-orology, vol. 2019, pp. 1–10, Article ID 2385310, 2019.

[8] M. Konda, N. Imasato, and A. Shibata, “A new method todetermine near-sea surface air temperature by using satellitedata,” Journal of Geophysical Research Oceans, vol. 101, no. C6,pp. 14349–14360, 1996.

[9] U. Marcel, “Comparison of satellite-derived land surfacetemperature and air temperature frommeteorological stationson the pan-arctic scale,” Remote Sensing, vol. 5, no. 5,pp. 2348–2367, 2013.

[10] W. Wagner, V. Naeimi, K. Scipal, R. de Jeu, and J. Mart́ınez-Fernández, “Soil moisture from operational meteorologicalsatellites,” Hydrogeology Journal, vol. 15, no. 1, pp. 121–131,2007.

[11] C. Oppenheimer, “Review article: volcanological applicationsof meteorological satellites,” International Journal of RemoteSensing, vol. 19, no. 15, pp. 2829–2864, 1998.

[12] H. W. Yates, J. D. Tarpley, S. R. Schneider, D. F. McGinnis,and R. A. Scofield, “'e role of meteorological satellites inagricultural remote sensing,” Remote sensingof environment,vol. 14, no. 1–3, pp. 219–233, 1984.

[13] C. O. Justice, “'e moderate resolution imaging spectro-radiometer (MODIS): land remote sensing for global changeresearch,” IEEE Transactions on Geoscience & Remote Sensing,vol. 36, no. 4, pp. 1228–1249, 2002.

[14] D. A. Chu, “Remote sensing of smoke from MODIS airbornesimulator during the SCAR-B experiment,” Journal of Geo-physical Research Atmospheres, vol. 103, no. D24, Article ID31979, 1998.

[15] L. Mohammadi, N. Molanian, and A. Heidari, “Determina-tion of the best coverage area for receiver stations of LEOremote sensing satellites,” in Proceedings of the 3rd Interna-tional Conference on Information and CommunicationTechnologies: From

framework for modelling and forecasting time series,”Knowledge-Based Systems, vol. 193, Article ID 105476, 2020.

[34] R. Bycroft, J. X. Leon, and D. Schoeman, “Comparing randomforests and convoluted neural networks for mapping ghostcrab burrows using imagery from an unmanned aerial ve-hicle,” Estuarine, Coastal and Shelf Science, vol. 224, pp. 84–93, 2019.

[35] Y. Lee, D. Han, M.-H. Ahn, J. Im, and S. J. Lee, “Retrieval oftotal precipitable water from himawari-8 AHI data: A com-parison of random forest, extreme gradient boosting, anddeep neural network,” Remote Sensing, vol. 11, no. 15, p. 1741,2019.

[36] J. Garćıa-Gutiérrez, F. Mart́ınez-Álvarez, A. Troncoso, andJ. C. Riquelme, “A comparison of machine learning regressiontechniques for LiDAR-derived estimation of forest variables,”Neurocomputing, vol. 167, no. 1, pp. 24–31, 2015.

[37] D. Upreti,W. Huang,W. Kong et al., “A comparison of hybridmachine learning algorithms for the retrieval of wheat bio-physical variables from sentinel-2,” Remote Sensing, vol. 11,no. 5, p. 481, 2019.

[38] R. Li, L. Cui, H. Fu, Y. Meng, J. Li, and J. Guo, “Estimatinghigh-resolution PM1 concentration from Himawari-8 com-bining extreme gradient boosting-geographically and tem-porally weighted regression (XGBoost-GTWR),” AtmosphericEnvironment, vol. 229, Article ID 117434, 2020.

[39] G. Papacharalampous and H. Tyralis, “Hydrological timeseries forecasting using simple combinations: Big data testingand investigations on one-year ahead river flow predictabil-ity,” Journal of Hydrology, vol. 590, Article ID 125205, 2020.

[40] R. Pérez-Chacón, G. Asencio-Cortés, F. Mart́ınez-Álvarez,and A. Troncoso, “Big data time series forecasting based onpattern sequence similarity and its application to the elec-tricity demand,” Information Sciences, vol. 540, pp. 160–174,2020.

[41] H. Abdollahi, “A novel hybrid model for forecasting crude oilprice based on time series decomposition,” Applied Energy,vol. 267, Article ID 115035, 2020.

[42] R. S. dos Santos, “Estimating spatio-temporal air temperaturein London (UK) using machine learning and earth obser-vation satellite data,” International Journal of Applied EarthObservation and Geoinformation, vol. 88, Article ID 102066,2020.

[43] H. J. Richardson, D. J. Hill, D. R. Denesiuk, and L. H. Fraser,“A comparison of geographic datasets and fieldmeasurementsto model soil carbon using random forests and stepwise re-gressions (British Columbia, Canada),” GI Science & RemoteSensing, vol. 54, no. 4, pp. 573–591, 2017.

[44] H. Franco-Lopez, A. R. Ek, and M. E. Bauer, “Estimation andmapping of forest stand density, volume, and cover type usingthe k-nearest neighbors method,” Remote Sensing of Envi-ronment, vol. 77, no. 3, pp. 251–274, 2001.

[45] R. Haapanen, A. R. Ek, M. E. Bauer, and A. O. Finley,“Delineation of forest/nonforest land use classes using nearestneighbor methods,” Remote Sensing of Environment, vol. 89,no. 3, pp. 265–271, 2004.

[46] M. Belgiu and L. Drăguţ, “Random forest in remote sensing: areview of applications and future directions,” ISPRS Journal ofPhotogrammetry and Remote Sensing, vol. 114, no. 114,pp. 24–31, 2016.

[47] M. J. Cracknell and A.M. Reading, “'e upside of uncertainty:Identification of lithology contact zones from airborne geo-physics and satellite data using random forests and supportvector machines,” Geophysics, vol. 78, no. 3, 2013.

[48] X. Ye, X. Yang, X. Xiong, Y. Shen, M. Hao, and R. Gu, “Aquality control method based on an improved random forestalgorithm for surface air temperature observations,” Advancesin Meteorology, vol. 2017, pp. 1–15, Article ID 8601296, 2017.

[49] B. Babar, L. T. Luppino, T. Boström, and S. N. Anfinsen,“Random forest regression for improved mapping of solarirradiance at high latitudes,” Solar Energy, vol. 198, pp. 81–92,2020.

[50] L. V. Utkin, M. S. Kovalev, and F. P. A. Coolen, “Impreciseweighted extensions of random forests for classification andregression,” Applied Soft Computing, vol. 92, Article ID106324, 2020.

[51] L. Liu, M. Ji, and M. Buchroithner, “Combining partial leastsquares and the gradient-boosting method for soil propertyretrieval using visible near-infrared shortwave infraredspectra,” Remote Sensing, vol. 9, no. 12, p. 1299, 2017.

[52] J. Son, I. Jung, K. Park, and B. Han, “Tracking-by-Segmen-tation with online gradient boosting decision tree,” in Pro-ceedings of the International Conference on Computer Vision,Las Condes, Chile, December 2015.

[53] H. Mo, H. Sun, J. Liu, and S. Wei, “Developing windowbehavior models for residential buildings using XGBoostalgorithm,” Energy and Buildings, vol. 205, no. 15, Article ID109564, 2019.

[54] M. H. D. M. Ribeiro and L. dos Santos Coelho, “Ensembleapproach based on bagging, boosting and stacking for short-term prediction in agribusiness time series,” Applied SoftComputing, vol. 86, Article ID 105837, 2020.

[55] I. B. Mustapha and F. Saeed, “Bioactive molecule predictionusing extreme gradient boosting,” Molecules, vol. 21, no. 8,p. 983, 2016.

[56] C. Li, “Power load forecasting based on the combined modelof LSTM and XGBoost,” in Proceedings of the InternationalConference on Pattern Recognition, Wenzhou, China, June2019.

[57] Z.-Y. Chen, T.-H. Zhang, R. Zhang et al., “Extreme gradientboosting model to estimate PM2.5 concentrations withmissing-filled satellite data in China,” Atmospheric Environ-ment, vol. 202, pp. 180–189, 2019.

[58] S. Zhao, D. Zeng, W. Wang et al., “Mutation grey wolf elitePSO balanced XGBoost for radar emitter individual identi-fication based on measured signals,” Measurement, vol. 159,Article ID 107777, 2020.

[59] S. J. Lee, M.-H. Ahn, and Y. Lee, “Application of an artificialneural network for a direct estimation of atmospheric in-stability from a next-generation imager,” Advances in At-mospheric Sciences, vol. 33, no. 2, pp. 221–232, 2016.

[60] J. Tang, C. Deng, G.-B. Huang, and B. Zhao, “Compressed-domain ship detection on spaceborne optical image usingdeep neural network and extreme learning machine,” IEEETransactions on Geoscience and Remote Sensing, vol. 53, no. 3,pp. 1174–1185, 2015.

[61] G. T. Ribeiro, V. C. Mariani, and L. Coelho, “Enhancedensemble structures using wavelet neural networks applied toshort-term load forecasting,” Engineering Applications ofArtificial Intelligence, vol. 82, pp. 272–281, 2019.

[62] W. Wang, X. Sun, R. Zhang, Z. Li, Z. Zhu, and H. Su, “Multi-layer perceptron neural network based algorithm for esti-mating precipitable water vapour from MODIS NIR data,”International Journal of Remote Sensing, vol. 27, no. 3,pp. 617–621, 2006.

[63] K. I. Chronopoulos, I. X. Tsiros, I. F. Dimopoulos, andN. Alvertos, “An application of artificial neural networkmodels to estimate air temperature data in areas with sparse

Advances in Meteorology 13

network of meteorological stations,” Journal of EnvironmentalScience and Health, Part A, vol. 43, no. 14, pp. 1752–1757,2008.

[64] S. Rodrigues Moreno, R. Gomes da Silva, V. Cocco Mariani,and L. dos Santos Coelho, “Multi-step wind speed forecastingbased on hybrid multi-stage decomposition model and longshort-term memory neural network,” Energy Conversion andManagement, vol. 213, Article ID 112869, 2020.

[65] A. Bayat and S. Mashhadizadeh Maleki, “Comparison ofprecipitable water vapor derived from AIRS and SPM mea-surements and its correlation with surface temperature of 29synoptic stations over Iran,” Journal of Atmospheric and Solar-Terrestrial Physics, vol. 178, pp. 24–31, 2018.

[66] W. W. Gong, “Evaluation of surface meteorological elementsfrom several numerical models in China,” Climatic and En-vironmental Research, vol. 20, no. 1, pp. 53–62, 2015.

[67] D. Pozo Vázquez, F. J. Olmo Reyes, and L. Alados Arboledas,“A comparative study of algorithms for estimating landsurface temperature from AVHRR Data,” Remote Sensing ofEnvironment, vol. 62, no. 3, pp. 215–222, 1997.

14 Advances in Meteorology