Search

Sentinel-2 satellites provide dense image time series exhibiting high spectral, spatial and temporal resolutions. These images are in particular of utter interest for Land-Cover (LC) mapping at large scales. LC maps can now be computed on a yearly basis at the scale of a country with efficient supervised classifiers, assuming suitable training data are available. However, the efficient exploitation of large amount of Sentinel-2 imagery still remain challenging on unexplored areas where state-of-the-art classifiers are prone to fail. This paper focuses on Land-Cover mapping over Cameroon for the purpose of updating the Very High Resolution national topographic geodatabase. The ι2 framework is adopted and tested for the specificity of the country. Here, experiments focus on generic vegetation classes (five) which enables providing robust focusing masks for higher resolution classifications. Two strategies are compared: (i) a LC map is calculated out of a year long time series and (ii) monthly LC maps are generated and merged into a single yearly map. Satisfactory accuracy scores are obtained (>94% in Overall Accuracy), allowing to provide a first step towards finer-grained map retrieval. [ABSTRACT FROM AUTHOR]

Sentinel-2 satellites provide dense image time series exhibiting high spectral, spatial and temporal resolution. These images are in particular of utter interest to map Land-Cover (LC) at large scale. LC maps can now be computed on a yearly basis at the scale of a country with efficient supervised classifiers, assuming suitable training data are available. However, the efficient exploitation of large amount of Sentinel-2 imagery still remain challenging on unexplored areas where state-of-the-art classifiers are prone to fail. This paper focuses on Land-Cover mapping over Cameroon for the purpose of updating the national topographic geodatabase. The ι2 framework is adopted and tested for the specificity of the country. Here, experiments focus on generic classes (five) which enables providing robust focusing masks for higher resolution classifications. Two strategies are compared: (i) a LC map is calculated out of a year long time series and (ii) monthly LC maps are generated and merged into a single yearly map. Satisfactory accuracy scores are obtained, allowing to provide a first step towards finer-grained map retrieval. [ABSTRACT FROM AUTHOR]