| The accuracy of forest species composition classification using high-resolution imagery can be increased using image textures. However, a challenge in remote sensing is how to represent the classification results to end-users. Kriging analysis, based on the image semivariogram measured the spatial dependence of pixels, estimates local pixel value by the values of neighbouring pixels, through which unknown scaled spatial variances (noise) are reduced. From a created kriging surface, image textures can be derived and used to improve the classification accuracy; the maps of classification are more readily interpreted and used by non-remote sensing forest specialists. With a combination of original spectral data, an NDVI kriging surface and image textural data from 2 m CASI imagery, an 86% of the classification accuracy with 0.82 of Khat was obtained. A low-pass filter (mode filtering, 7 x 7) was applied for the final classification of forest species composition, which closely resembled a map produced from aerial photo interpretation techniques in the GIS database. |
