9af88626-f791-4b61-ad49-b18e53641d6f Olivier Flores Olivier Flores Olivier Flores University of Reunion, PVBMT University of Reunion, PVBMT Université de La Réunion, PVBMT olivier.flores@univ-reunion.fr olivier.flores@univ-reunion.fr olivier.flores@univ-reunion.fr Victor K/BIDI Victor K/BIDI Victor K/BIDI OSU-Réunion, UAR3365 OSU-Réunion, UAR3365 OSU-Réunion, UAR3365 Data steward Data steward +262 262 93 82 28 15 avenue René Cassin, Université de la Réunion 15 avenue René Cassin, Université de la Réunion Saint-Denis 97744 France France victor.kbidi@univ-reunion.fr victor.kbidi@univ-reunion.fr victor.kbidi@univ-reunion.fr 2026-01-19T11:30:57 ISO 19115:2003/19139 1.0 WGS 1984 WGS 1984 Land cover maps of La Réunion islands in 1950 and 2022 Land cover maps of La Réunion islands in 1950 and 2022 Cartes de couverture des sols de l’île de La Réunion en 1950 et 2022 2026-01-12 The resource provides two land cover maps of Réunion Island for the years 1950 and 2022 derived from the analysis of ortho-photographs at the island scale (Source IGN). The produced typology uses five cover classes: forest, low vegetation, agriculture, urban, and shadow (related to topography). The method used is based on encoding the two aligned rasters, converted into a single band of grayscale for 2022, using a vision-transformer deep learning model. From the features calculated for each pixel, a random forest classification model is trained separately for each year using a set of ROIs (Regions Of Interest), target polygons delineated within each of the selected classes through photo-interpretation of the original images. Model validation is performed on independent sets of polygons also defined by photo-interpretation. The maps provided in the resource are derived from the prediction of cover classes for both years using the trained and validated models. These are raw predictions, meaning that no post-processing has been applied to reduce potential noise due to classification errors. The shared resource is part of the results from the FRAG'ILE research and development program (FRAGmentation en milieu InsuLairRE, UR / CBNM/ IRD, funded by OFB, https://fragile.frama.io). The resource provides two land cover maps of Réunion Island for the years 1950 and 2022 derived from the analysis of ortho-photographs at the island scale (Source IGN). The produced typology uses five cover classes: forest, low vegetation, agriculture, urban, and shadow (related to topography). The method used is based on encoding the two aligned rasters, converted into a single band of grayscale for 2022, using a vision-transformer deep learning model. From the features calculated for each pixel, a random forest classification model is trained separately for each year using a set of ROIs (Regions Of Interest), target polygons delineated within each of the selected classes through photo-interpretation of the original images. Model validation is performed on independent sets of polygons also defined by photo-interpretation. The maps provided in the resource are derived from the prediction of cover classes for both years using the trained and validated models. These are raw predictions, meaning that no post-processing has been applied to reduce potential noise due to classification errors. The shared resource is part of the results from the FRAG'ILE research and development program (FRAGmentation en milieu InsuLairRE, UR / CBNM/ IRD, funded by OFB, https://fragile.frama.io). La ressource met à disposition deux cartes de couverture des sols de La Réunion pour les années 1950 et 2022 issues de l’analyse des ortho-photographies à l’échelle de l’île (Source IGN). La typologie produite utilise cinq classes de couvert : forêt, végétation basse, agriculture, urbain et ombre (liée au relief). La méthode utilisée se base sur l’encodage des deux rasters alignés, et converti en une seul bande de nuances de gris pour 2022, à partir d’une modèle d’apprentissage profond (deep learning) de type vision-transformer. A partir des attributs (features) calculés pour chaque pixel, un modèle de classification de type random forest est entraîné séparément pour chaque année à partir d’un ensemble de ROI (Region Of Interest), polygones cibles délimités dans chacune des classes retenues à partir d’une photo-interprétation des images initiales. La validation des modèles se fait sur des ensembles indépendants de polygones également définis par photo-interprétation. Les cartes fournies dans la ressource sont issues de la prédiction des classes de couvert pour les deux années à partir des modèles entraînés et validés. Il s’agit des prédictions brutes, au sens où aucun post-traitement n’a été appliqué, pour réduire le bruit éventuel dû aux erreurs de classification. La ressource partagée fait partie des résultats du programme de recherche développement FRAG’ILE (FRAGmentation en milieu InsuLairRE, UR / CBNM/ IRD, financ. OFB, https://fragile.frama.io). To provide a historical land cover map of Réunion Island in 1950 and a recent land cover map in 2022. Since the maps were obtained using two similar approaches, their comparison allows for the assessment of the spatio-temporal dynamics of the five selected classes (forest, low vegetation, agriculture, urban, shadow). To provide a historical land cover map of Réunion Island in 1950 and a recent land cover map in 2022. Since the maps were obtained using two similar approaches, their comparison allows for the assessment of the spatio-temporal dynamics of the five selected classes (forest, low vegetation, agriculture, urban, shadow). Fournir une cartographie de l’occupation du sol historique de l’île de La Réunion en 1950 et une cartographie de l’occupation récente en 2022. Les cartes ayant été obtenues selon deux approches analogues, leur comparaison permet d’évaluer la dynamique spatio-temporelle des cinq classes retenues (forêt, végétation basse, agriculture, urbain, ombre). Anthony Malkassian Anthony Malkassian Anthony Malkassian University of Reunion, PVBMT University of Reunion, PVBMT Université de La Réunion, PVBMT anthony.malkassian@univ-reunion.fr anthony.malkassian@univ-reunion.fr anthony.malkassian@univ-reunion.fr Olivier Flores Olivier Flores Olivier Flores University of Reunion, PVBMT University of Reunion, PVBMT Université de La Réunion, PVBMT olivier.flores@univ-reunion.fr olivier.flores@univ-reunion.fr olivier.flores@univ-reunion.fr Robin Pouteau Robin Pouteau Robin Pouteau IRD IRD IRD robin.pouteau@ird.fr robin.pouteau@ird.fr robin.pouteau@ird.fr https://geosur.osureunion.fr:/geonetwork/srv/api/records/9af88626-f791-4b61-ad49-b18e53641d6f/attachments/fragile_lcc_forest_1950-2022.jpg Land cover Land cover Couverture du sol Historical aerial imagery Historical aerial imagery Images aériennes historiques Classification Classification Classification Deep learning Deep learning Deep learning License CC BY NC SA Specify the citation format : Malkassian A., Pouteau R., Flores O., Land cover maps of La Réunion islands in 1950 and 2022, https://doi.org/10.26171/a3jm-9c40 License CC BY NC SA Specify the citation format : Malkassian A., Pouteau R., Flores O., Land cover maps of La Réunion islands in 1950 and 2022, https://doi.org/10.26171/a3jm-9c40 License CC BY NC SA Indiquer le format de citation : Malkassian A., Pouteau R., Flores O., Land cover maps of La Réunion islands in 1950 and 2022, https://doi.org/10.26171/a3jm-9c40 environment biota 1950-01-01 2012-12-31 55.1331 55.8947 -21.4096 -20.8175 https://doi.org/10.26171/a3jm-9c40 DOI DOI DOI DOI