An object-conditional approach for satellite remote sensing land cover mapping in African Savannas

Abstract

To understand human and environmental interactions such as climate change and as land cover (LC) change drivers, a standardized, updatable and functional land cover mapping mechanisms is imperative. Moreover, LC data is important for cellular automation (CA) and multi-agents models (MASs) in order to predict current land use, land feature dynamics and future land use changes. However, transparent LC semantics in a standardized ‘language’ throughout Africa is still missing. Also, many LC data initiatives render ambiguous and rigid LC definitions, without consideration of LC transitions zones, LC modifiers or class functionality/comparability. Seemingly there is a need to conceptualize LC mapping mechanisms in such a way that the temporal and spatial dynamics are well described and can be readily updated once a new and more innovative LC mapping method becomes available. One solution is to utilize the FAO standard Land Cover Classification (LCCS) coding system, which does not define rigid LC polygons boundaries but is rather based on vegetation life form separation. Also, life form attributes can be attached on several levels of details according to the requirements for efficient LC mapping. Within the GLOWA Volta project in West Africa, we tested an innovative object-conditional approach to map LC. We used 30-Landsat ‘snapshots’, 26 Landsat tiles from the year 1990 and from 2000/2001 correspondingly, as well as 250-MODIS time-series data sets from 2000 to 2005. The conditions were set by using the LCCS legend, and object-behavior was considered by using only spectral wavebands and information layers relevant to a certain LC object, i.e. the spatial and temporal behavior of the object. In addition, topographical variables from 90-meter resolution SRTM digital elevation model were used. The occurrence of woody and herbaceous life forms, as well as managed herbaceous areas and water spaces were mapped using a spectral statistical neighborhood fuzzy convolution filter on the Landsat data. Each potential LC feature was assessed according to its statistical significance, considering all spectral bands as a function of the significances of the neighboring pixels. Initial results show that highly dynamic LC features such as wetlands can be most feasible mapped using dry season MODIS NDVI time-series data sets (cumulatively). Wet season Near-IR reflectances from MODIS could be further used to characterize wetland flooding levels, in conjunction with the SRTM DEM derived landscape positions and streamline flow patterns. Using the fuzzy convolution filter on the Landsat tiles, the highest accuracies were attained in mapping woody life form with different tree cover densities (75% error of commission; 82% error of omission). The MODIS wetlands were mapped with the second highest certainty (69% commission error and 66% omission error). Areas with open or sparse woody vegetation as well as savanna spaces were the most difficult classes to correctly map (error of commissions of 45 %). We conclude that the object-conditional approach shows promising results. The LCCS legend codes are flexible to allow for regular LC updates and feasible class discriminations on different levels of detail and using multiple data sets. The LCCS data is also compatible to other projects in Africa that are currently also using LCCS to map land cover.