Improving the accuracy of extant gridded population maps using multisource map fusion

Figures & data

Figure 1. Location of the study area and population distribution of the township.

Table 1. Input datasets for multisource map fusion

Name	Year(s)	Spatial resolution	Address
GPW4	2015	30 arc-seconds	https://sedac.ciesin.columbia.edu/data/collection/gpw-v4/sets/browse
LandScan	2015	30 arc-seconds	https://landscan.ornl.gov/
WorldPop	2015	30 arc-seconds	https://www.worldpop.org/?tdsourcetag=s_pctim_aiomsg
Population sample survey data	2015	Township	https://data.cnki.net/
Demographic data	2005–2015	Village	http://fz.wanfangdata.com.cn/index.do
Impervious surfaces data	2015	30 m	http://data.ess.tsinghua.edu.cn/
Mobile phone positioning data	2015	0.01 decimal degree	https://heat.qq.com/

Figure 2. Distribution maps of the gridded population in the study area: (A) BTRG, (B) Guizhou Province, (C) Guangdong Province; (a) GPW4, (b) LandScan, and (c) WorldPop.

Figure 3. Approaches adopted for producing reference data for an evenly populated area. (a) Impervious surfaces at 30 m resolution, (b) village coordinates, (c) Tyson polygons, (d) Tyson polygons for which the population density information was obtained, (e) impervious surfaces with population density obtained after grid partitioning, and (f) reference data for the area with an even population distribution.

Figure 4. Framework for achieving multisource map fusion.

Table 2. Percentage of townships for multiple input maps with different relative error ranges

Area	Name	Relative error
		0–10%	10%-30%	30%-50%	Total
BTHR	GPW4	22.45%	42.32%	16.39%	81.16%
	LandScan	16.13%	30.32%	24.00%	70.45%
	WorldPop	26.58%	43.61%	16.13%	86.32%
Guangdong Province	GPW4	24.01%	37.93%	18.69%	80.63%
	LandScan	15.83%	30.97%	21.56%	68.35%
	WorldPop	24.69%	41.75%	17.05%	83.49%
Guizhou Province	GPW4	30.53%	43.51%	18.32%	92.37%
	LandScan	14.12%	25.95%	26.34%	66.41%
	WorldPop	30.15%	46.18%	18.70%	95.04%

Note. Input the number of townships within different relative error intervals on the map as a proportion of the total number of townships in the sparsely populated area.

Figure 5. Heat map representing the root-mean-squared error of different algorithms (a for RF algorithm, b for VW algorithm, c for GR algorithm; A refers to Guizhou Province, B is Guangdong Province, and C is BTHR).

Note. The VW algorithm refers to the variance weighted algorithm; GR algorithm is the Granger-Ramanathan algorithm; RF algorithm denotes the random forest algorithm.

Figure 6. Heat map representing the mean absolute error of different algorithms (a for RF algorithm, b for VW algorithm, c for GR algorithm; A refers to Guizhou Province, B is Guangdong Province, and C is BTHR).

Note. The VW algorithm refers to the variance weighted algorithm; GR algorithm is the Granger-Ramanathan algorithm; RF algorithm denotes the random forest algorithm.

Table 3. Evaluation indicator values reported for different data in a total perspective

Area	BTHR			Guangdong		Guizhou
Name	RMSE		MAE	RMSE	MAE	RMSE	MAE
GPW4	2980.35	1098.92		5700.71	2194.98	2402.07	651.09
LandScan	3287.09	1257.21		6296.72	2443.01	2282.40	711.22
WorldPop	2859.71	1064.21		5330.00	1917.32	2114.98	542.53
VW algorithm	2469.12	958.43		4810.89	1816.08	2008.07	562.70
GR algorithm Case 1	2895.09	1197.84		5593.11	2351.01	2498.51	672.42
RF algorithm Case 1	2841.56	1057.99		5275.79	2093.17	2300.35	616.08
GR algorithm Case 2	2701.06	1005.27		5070.48	1829.90	2221.68	564.61
RF algorithm Case 2	2569.67	847.17		4505.62	1566.58	2130.30	509.14

Note. RMSE refers to root mean squared error; MAE is mean absolute error; VW algorithm refers to the variance weighted algorithm; GR algorithm is the Granger-Ramanathan algorithm; RF algorithm denotes the random forest algorithm; Case 1 is map fusion for the input maps only; Case 2 adds two ancillary variables to Case 1.

Figure 7. Taylor diagram describing the overall accuracy of the different data. Note. A: GPW4, B: LandScan, C: WorldPop, D: Variance weighted algorithm, E: Granger-Ramanathan algorithm with Case 1, F: Random forest algorithm with Case 1, G: Granger-Ramanathan algorithm with Case 2, H: Random forest algorithm with Case 2; a: BTHR, b: Guizhou Province, c: Guangdong Province.

Figure 8. Final output: The most accurate fusion result is spliced with WorldPop in the sparsely populated area (A: BTHR, B: Guizhou Province, C: Guangdong Province).

Table 4. Relative importance of different variables for the RF algorithm in BTHR, Guangdong Province, and Guizhou Province

Name	RF algorithm Case 1			RF algorithm Case 2
	BTHR	Guangdong	Guizhou	BTHR	Guangdong	Guizhou
GPW4	30%	27%	27%	14%	15%	20%
LandScan	33%	28%	33%	15%	16%	17%
WorldPop	38%	45%	40%	22%	24%	32%
ISP				6%	6%	4%
NP				42%	39%	27%

Note. The GR algorithm refers to the Granger-Ramanathan algorithm; the RF algorithm denotes the random forest algorithm; ISP is the impervious surface percentage; and NP represents the number of mobile phone positionings. The numbers in the table indicate the importance of the variables.

Data availability statement

The data and codes that support the findings of this study are openly available in [figshare] at [https://www.doi.org/10.6084/m9.figshare.14740272].