High-resolution land cover classification: cost-effective approach for extraction of reliable training data from existing land cover datasets

Figures & data

Figure 1. Region of interest.

Figure 2. Illustration of the derivation of the Map Of Land Cover Agreement – MOLCA.

Figure 3. Overview of MOLCA production workflow.

Table 1. OA of land cover products with different validation datasets.

	Validation dataset
Classification test	All areas	Easy-to-classify	Hard-to-classify
Planet 4B	73%	88%	49%
Sentinel-2 4B	68%	84%	41%
Sentinel-2 allB	78%	93%	52%

Table 2. Accuracy of Sentinel-2 allB product with all samples.

	Bareland	Built-up	Cropland	Forest	Grassland	Ice and snow	Shrubland	Water	UA	PA	OA
Bareland										0%	78%
Built-up		213	7	2	14		6		88%	98%
Cropland		1	99	18	36		13		59%	60%
Forest		1	17	282	29	2	68		71%	88%
Grassland	2	2	37	13	190		66	1	61%	65%
Ice and snow										0%
Shrubland	1	1	5	2	25		311		90%	67%
Water			1	4			1	205	97%	100%

Table 3. Accuracy of Sentinel-2 allB product with hard-to-classify samples.

	Built-up	Cropland	Forest	Grassland	Ice and snow	Shrubland	Water	UA	PA	OA
Built-up	1	4	2	11		6		4%	100%	52%
Cropland		34	11	12		11		50%	38%
Forest		15	99	27	2	68		47%	76%
Grassland		31	13	52		60	1	33%	43%
Ice and snow									0%
Shrubland		5	1	18		123		84%	46%
Water		1	4			1	19	76%	95%

Table 4. Accuracy of Sentinel-2 allB product with easy-to-classify samples.

	Bareland	Built-up	Cropland	Forest	Grassland	Shrubland	Water	UA	PA	OA
Bareland									0%	93%
Built-up		212	3		3			97%	98%
Cropland		1	65	7	24	2		66%	86%
Forest		1	2	183	2			97%	96%
Grassland	2	2	6		138	6		90%	79%
Shrubland	1	1		1	7	188		95%	96%
Water							186	100%	100%

Figure 4. Land cover map obtained in Sentinel-2 allB classification test.

Table 5. Comparison of the accuracy of World Cover, Dynamic World and Esri land cover with the accuracy of Sentinel-2 allB product.

	World Cover		Dynamic World		Esri land cover		Sentinel-2 allB
	PA	UA	PA	UA	PA	UA	PA	UA
Bareland	0%	0%	33%	6%	0%	0%		0%
Built-up	80%	98%	98%	88%	98%	93%	88%	98%
Cropland	6%	85%	9%	20%	4%	60%	59%	60%
Forest	94%	60%	95%	43%	95%	41%	71%	88%
Grassland	63%	56%	27%	72%	51%	80%	61%	65%
Shrubland	72%	75%	57%	81%			90%	67%
Water	98%	99%	93%	97%	100%	97%	97%	100%
Wetland	0%	0%
Ice and snow	50%	100%	100%	40%				0%
OA	72%		64%		67%		78%

Data availability statement

The datasets that support this study are freely available land cover datasets such as: FROM-GLC (http://data.ess.tsinghua.edu.cn/fromglc2017v1.html), GL30 (http://www.globallandcover.com/), CCI Africa Prototype (https://2016africalandcover20m.esrin.esa.int/), GHS BU S1NODSM (http://cidportal.jrc.ec.europa.eu/ftp/jrc-opendata/GHSL/GHS_BUILT_S1NODSM_GLOBE_R2018A/), WSF (https://download.geoservice.dlr.de/WSF2019/files/), GSW (https://global-surface-water.appspot.com/download), and FNF (https://www.eorc.jaxa.jp/ALOS/en/palsar_fnf/data/index.htm). GL30 and FNF require registration in order to retrieve the data, while CCI Africa Prototype require acceptance of terms of use. Satellite imagery, Planet NICFI Basemap (https://developers.google.com/earth-engine/datasets/catalog/projects_planet-nicfi_assets_basemaps_africa?hl=en) and Sentinel-2 (https://developers.google.com/earth-engine/datasets/catalog/COPERNICUS_S2), as well as CGIAR-CSI SRTM DEM (https://developers.google.com/earth-engine/datasets/catalog/CGIAR_SRTM90_V4), World Cover (https://developers.google.com/earth-engine/datasets/catalog/ESA_WorldCover_v100), Dynamic World (https://developers.google.com/earth-engine/datasets/catalog/GOOGLE_DYNAMICWORLD_V1) were obtained from Earth Engine Data Catalog. Esri land cover is directly accessible on Google Earth Engine through ‘awesome-gee-community-catalog’ by using ID projects/sat-io/open-datasets/landcover/ESRI$\mathrm{\_}$Global-LULC$\mathrm{\_}$10m$\mathrm{\_}$TS. ‘awesome-gee-community-catalog’ is a collection of geospatial datasets contributed by the community. MOLCA, which is the first dataset obtained by the methodology reported in this paper, is freely and openly available at https://doi.org/10.5281/zenodo.8071675.