Factors affecting the accuracy of forest clear-cut area estimation on medium spatial resolution satellite winter images

The aim of this paper was to investigate the influence of attributes, describing clear-cut patch size, age, shape, nearest neighbours and habitat conditions on classification results of medium spatial resolution Landsat Thematic Mapper (TM) satellite images. The study area was Võru State Forest District in South-East Estonia and the satellite images used were made in late winter when the ground was covered with snow. Clear-cuts as significantly changed areas in forests were discerned from non-changed areas with thresholding of a two-date difference image. The results of the statistical analysis showed that clear-cut attributes had an influence on image classification results. The most influential variables (presented in decreasing order of significance) are the mean of clear-cut patch pixel values on the difference image, the relative boundary length with forest, the relative boundary length with coniferous forest and the clear-cut area to perimeter ratio. The age of clear-cut and habitat conditions had no statistically significant influence on classification results. The set of influential attributes remained the same when the classifications were performed on two more liberal and on two more conservative thresholding levels. Images in the visible and near infrared spectral region (Landsat TM bands 1-4) revealed appropriate for clear-cut mapping. The difference in the area of a single clear-cut patch represented in the forestry database to that classified from a Landsat TM image was about a sixth of the patch size. This implies the utility of medium resolution satellite images in clear-cut activity assessments in particular areas but not so much the applicability of these images for single patch area estimations.