Satellite remotely sensed data: potential for mapping Uganda's savannas

Abstract

The sustainable use and management of Uganda's savannas poses a big challenge because of, inter-alia, the lack of accurate, cost-effective and up-to-date spatial information. In the present study, it was postulated that cost effective remote sensing techniques, based on low/coarse resolution imagery and automated standard image classification techniques, are not optimized for the identification and mapping of land cover classes characteristic of wood/grass mixtures of different densities - savannas. The objectives of the study were to establish (l) the level of accuracy of existing land cover maps derived from low/coarse resolution imagery for Uganda's savannas; (2) the optimal spatial resolution for mapping Uganda's savannas; and (3) whether wood density can be harnessed to improve the accuracy of land corer maps generated from Landsat TM for Uganda's savannas. The study was carried out in four test sites located in Arua, Gulu, Nebbi and Masaka Districts of Uganda. The level of accuracy of each existing land cover map, for Uganda's savannas, was assessed using a region-based technique. The region-based technique is essentially a GIS cross-tabulation of two thematic maps, A and B, where A represented reference maps (derived from high-resolution IKONOS imagery) and B an existing land cover map derived from low/coarse resolution imagery. To evaluate the optimal image resolution, suitable for mapping Uganda's savannas, a triangulation of techniques was made. These techniques involved measuring trends of several parameters (including overall classification accuracy, the level of terrain noise, land cover index, and average patch size) of land cover maps obtained from imagery simulated to different spatial resolutions i.e. 0.5 to 1.0 m, 1.5 m, 2.0 m, 2.5 m, 4.5 m, 5.0 m, 6.0 m, 7.0 m, 8.0 m, 9.0 m and 10.0 m. On the other hand, the relationship between areal wood density and spectra of Landsat TM was determined by correlating spectral land cover classes (obtained from Landsat TM imagery) with wood density obtained from IKONOS imagery. In general, the study revealed that the accuracy of existing land cover maps, generated from low/coarse resolution imagery for Uganda's savannas, is unacceptably low (less than 50%). The coarser the image resolution, the lower the accuracy of the land cover maps generated for Uganda's savannas. The study revealed that a combination of the level of image noise (spectral variance) and the average patch size of savanna features is a more robust technique (than has been the case hitherto) for the determination of the optimal image resolution for mapping Uganda's savannas. Using the technique, the optimal resolution, suitable for mapping Uganda's savannas, was estimated to be 3 - 4 m. At an optimal spatial resolution range of 3 - 4 m, imagery (acquired for Uganda's savannas are characterized by insignificant image noise spectral variance) whilst at the same time preserving the geometric integrity of the smallest geographical features (individual trees). Another key finding of the present study was that wood/grass mixtures of density classes 15 - 30%, 30 - 50% and 50 - 80% are correlated to spectra of Landsat TM imagery if representing Uganda's savannas. It was demonstrated that the quantitative relationship between wood density and spectra of Landsat TM can be harnessed, using an appropriate napping framework, to improve the accuracy of land cover maps generated for Uganda's savannas. From the findings of the present study, it was concluded that existing land cover map, produced from low/coarse resolution imagery, are grossly misclassified and hence should not be used for applications that require accurate spatial information. To generate accurate land cover maps for Uganda's savannas, imagery acquired at an optimal resolution of 3.5 +/-0.5 m should be used. The harnessing of wood density, while using Landsat TM imagery with an appropriate mapping framework, has a potential to improve the accuracy of land cover classes of Uganda's savannas. However, further research is recommended to operationalise the use of wood density for mapping Uganda's savannas using Landsal TM, or equivalent imagery.