Land use land cover very high resolution map (1.5-m) for the area of Korhogo, Côte d'Ivoire, 2018

Use email button above to contact.

It contains 16 land cover classes: ligneous savanna, crop, marsh, riparian forest, dense forest, open forest, rice, cotton, fallows, cashew plantation, mango plantation, built-up, bare soil, main roads, permanent water bodies and running waters.

The method used to generate the map involved a supervised object-based image classification using multisource satellite products (SPOT 6, Sentinel-2, a Digital Elevation Model), a ground-truth dataset acquired by fieldwork and photo-interpretation, and a random forest classifier. The classification accuracy is 83%.

In addition to the LULC georeferenced raster data, we propose the following files in this release:

• the raster attribute table, including definitions of the land cover classes in English and French ;
• a layer style (*.qml) to visualize the raster in QGIS ;
• the map as a .jpeg image (for visualization purposes only) ;
• representative pictures of the land cover classes ;
• the detailed methodology used to generate the data, with the resulting confusion matrix (in English and French) ;
• the ground-truth georeferenced dataset used for training and test in the classification ;
• the R script used to create the product, based exclusively on Free and Open Source software.

The satellite products used as input were: SPOT 6 images, Sentinel-2 images, and a Digital Elevation Model (DEM). We conducted a field campaign in December 2018 to build the ground-truth dataset. We established the land use classes in each of the zones based on bibliographic research on the types of landscapes potentially encountered in our zones and our observations of the landscape in situ. We collected a minimum of 20 plots per class, trying to spread them as much as possible over the extent of each zone. In the image processing step, we used the Baatz & Schape segmentation algorithm. We computed a total of about 100 predictors based on the SPOT 6, Sentinel-2 and DEM products, as well as the shape of the objects. We then trained a random forest model on the training data set. We generated the confusion matrix using the internal random forest validation procedure (based on out-of-bag observations). Based on this matrix, we then grouped, in the training dataset, land use classes with significant confusion (e.g. millet and sorghum growing areas). We trained a random forest model on this new version of the training dataset and then used it to predict the land cover class on each object from the segmentation. As before, we generated the confusion matrix and then extracted a classification quality index, the accuracy, measuring the proportion of correctly classified objects.

This work was supported by public funds received in the framework of GEOSUD, a project (ANR-10-EQPX-20) of the program "Investissements d’Avenir" managed by the French National Research Agency.