| _version_ | 1867171881000370176 |
|---|---|
| author | Phan, Thanh-Noi Dashpurev, Batnyambuu Wiemer, Felix Lehnert, Lukas |
| author_facet | Phan, Thanh-Noi Dashpurev, Batnyambuu Wiemer, Felix Lehnert, Lukas |
| collection | Datos científicos de ciencias marinas y ambientales |
| contents | The broad importance of land use and land cover information has been defined by and confirmed for many applications. Therefore, many land cover products have been developed at various scales (i.e., spatial resolution) and extensions (i.e., local, national, region, and global). Several studies have reported inconsistencies among global land cover (GLC) products causing that the accuracy of these products differ between regions. Recently, this issue has received a new level of attention, because many studies have pointed out that the inaccuracy of land cover products at regional scale can make a huge impact on the results of other applications relying on the GLC products (in the following downstream applications). Therefore, developing a method which can be easily and quickly applied to many different regions, but produce highly accurate land cover information is of utmost importance. To meet the first two criteria, several studies successfully used existing GLC products to automatically generate samples. However, none of these studies have been focused on the quality of the samples, which directly and largely affect the classification results. In this context, and taking Mongolia as a case study, we proposed a simple, fast, and accurate method to produce annual land cover maps with 250 m spatial resolution for entire Mongolia over a period of 20 years, from 2001 to 2020. The maps are based on MODIS data (products MOD13Q1 and MCD12Q1, version 6) and produced using modern machine learning techniques (the Random Forest) on the Google Earth Engine. Training samples have been selected by developing a semi-random approach which ensures that samples are spatially well-distributed, the number of samples for each class is in a similar order irrespective of the dominance of the land-cover classes and the samples are sufficiently apart from each other to reduce spatial autocorrelation. It is worth noting that we have selected Mongolia because of the low accuracy of GLC in this vast and remote country. Our results show that the accuracy of the new land cover maps improved compared with the corresponding MODIS products and if visually compared to Landsat images acquired at the same time. Overall accuracy from the validation data was approximately 90% for all new maps compared to 75% for the existing MODIS product. The result suggests that land cover maps, particularly for vegetation downstream application studies, can be largely improved based on the MODIS land cover products both regarding their spatial resolution and accuracy. Regarding Mongolia, these land cover maps are valuable e.g., for land degradation research, such as grassland monitoring, changes in forest cover, and monitoring desertification. Especially, information on grassland ecosystems is of utmost importance for Mongolia since more than half of the country economically depends on grassland resources. Therefore, Mongolia will benefit from the new dataset, not only economically, but also scientifically by helping researchers to discover more about the natural and social conditions of Mongolia. TIF file description: The land cover maps have 6 classes (land cover types), Water, Forest, Shrub land, Grassland, Others, and Bare land. Please note that water was masked out using the “JRC Global Surface Water Mapping Layers” (Pekel et al. 2016). LUC = Land use cover Spatial coverage: Mongolia (41.55709 - 52.17325 °N; 87.72279 - 120.04456°E) |
| format | Dataset Open Access |
| id | pangaea_https___doi_org_10_1594_PANGAEA_947514 |
| institution | PANGAEA |
| language | en |
| publishDate | 2022 |
| publisher | PANGAEA |
| record_format | pangaea |
| spellingShingle | Land cover classification maps of Mongolia from 2001 to 2020 Phan, Thanh-Noi Dashpurev, Batnyambuu Wiemer, Felix Lehnert, Lukas Google Earth Engine; Land cover; large scale; MODIS; Mongolia; MoreStep; MoreStep-Mobility at risk: Sustaining the Mongolian Steppe Ecosystem; MULT; Multiple investigations; Random Forest; Raster graphic, GeoTIFF format; Raster graphic, GeoTIFF format (File Size); semi-automatic The broad importance of land use and land cover information has been defined by and confirmed for many applications. Therefore, many land cover products have been developed at various scales (i.e., spatial resolution) and extensions (i.e., local, national, region, and global). Several studies have reported inconsistencies among global land cover (GLC) products causing that the accuracy of these products differ between regions. Recently, this issue has received a new level of attention, because many studies have pointed out that the inaccuracy of land cover products at regional scale can make a huge impact on the results of other applications relying on the GLC products (in the following downstream applications). Therefore, developing a method which can be easily and quickly applied to many different regions, but produce highly accurate land cover information is of utmost importance. To meet the first two criteria, several studies successfully used existing GLC products to automatically generate samples. However, none of these studies have been focused on the quality of the samples, which directly and largely affect the classification results. In this context, and taking Mongolia as a case study, we proposed a simple, fast, and accurate method to produce annual land cover maps with 250 m spatial resolution for entire Mongolia over a period of 20 years, from 2001 to 2020. The maps are based on MODIS data (products MOD13Q1 and MCD12Q1, version 6) and produced using modern machine learning techniques (the Random Forest) on the Google Earth Engine. Training samples have been selected by developing a semi-random approach which ensures that samples are spatially well-distributed, the number of samples for each class is in a similar order irrespective of the dominance of the land-cover classes and the samples are sufficiently apart from each other to reduce spatial autocorrelation. It is worth noting that we have selected Mongolia because of the low accuracy of GLC in this vast and remote country. Our results show that the accuracy of the new land cover maps improved compared with the corresponding MODIS products and if visually compared to Landsat images acquired at the same time. Overall accuracy from the validation data was approximately 90% for all new maps compared to 75% for the existing MODIS product. The result suggests that land cover maps, particularly for vegetation downstream application studies, can be largely improved based on the MODIS land cover products both regarding their spatial resolution and accuracy. Regarding Mongolia, these land cover maps are valuable e.g., for land degradation research, such as grassland monitoring, changes in forest cover, and monitoring desertification. Especially, information on grassland ecosystems is of utmost importance for Mongolia since more than half of the country economically depends on grassland resources. Therefore, Mongolia will benefit from the new dataset, not only economically, but also scientifically by helping researchers to discover more about the natural and social conditions of Mongolia. TIF file description: The land cover maps have 6 classes (land cover types), Water, Forest, Shrub land, Grassland, Others, and Bare land. Please note that water was masked out using the “JRC Global Surface Water Mapping Layers” (Pekel et al. 2016). LUC = Land use cover Spatial coverage: Mongolia (41.55709 - 52.17325 °N; 87.72279 - 120.04456°E) |
| title | Land cover classification maps of Mongolia from 2001 to 2020 |
| topic | Google Earth Engine; Land cover; large scale; MODIS; Mongolia; MoreStep; MoreStep-Mobility at risk: Sustaining the Mongolian Steppe Ecosystem; MULT; Multiple investigations; Random Forest; Raster graphic, GeoTIFF format; Raster graphic, GeoTIFF format (File Size); semi-automatic |
| url | https://doi.org/10.1594/PANGAEA.947514 |