Introduction
Classifications of land cover differ in how people define different land features. Differences in classifications of forests stem from the tree cover thresholds that people identify as constituting a forest. The Hansen global treemap is one that is widely used in research and studies across the world. This study assesses the Hansen tree cover dataset in Rwanda by comparing it to a custom-created classification of tree cover in Rwanda.
Analysis and Findings
My supervised classification using point data from the selected Landsat 8 image in Rwanda resulted in a classification that very closely mapped actual forest cover. The classification, shown in Figure 1, covered all existing forests and their boundaries. However, it also included scattered trees that may have may have been the dominant feature in that pixel (shown in Figure 2), which may be misleading to think of as a ‘forest’. The classification also included green fields (tilled and non-tilled) in certain areas where there were possible issues distinguishing between the spectral profiles of varying forest cover and green fields.
For initial comparison, I used the Hansen dataset where tree cover covered greater than (or equal to) 30% of the pixel area. This resulted in an image that included large swathes of green fields. The Hansen dataset’s definition of forest cover includes any tree that is taller than 5 meters and does not include any metrics for tree aggregation or clumping. Therefore, trees that lined or delineated green fields resulted in pixels showing forest cover is present. The agreement/disagreement map is shown in Figure 3 below along with the metrics chart in Figure 4.
The specific site chosen within Rwanda included a landscape that consisted mostly of dry fields and dark forest cover patches. Therefore, there was great agreement between the two images in what was not forest cover, showing high true negative rates, as demonstrated in Figure 4 in blue. True positive rates, shown in green, where both images agreed on where forest cover is, seemed to be less than the combined disagreement from the red and yellow categories. However, since this is a very specific subset of the dataset that seems to have more distinguished visual categories of landcover, this same agreement would not apply to all parts of the Landsat image.
For the purposes of this specific land selection, a balanced rate between true positive and true negative rates seem to be achieved best when the threshold of tree cover in the Hansen dataset is set to 35. This is where true positive and true negative rates are slightly higher than at a threshold of 30, as shown in Figure 5 below. After this value, the disagreement increases. From employing different threshold levels, it seems as though true positive rates are highest with lower thresholds, where tree cover in the Hansen dataset would be less specific or controlled and would cover most of my classified tree cover. At higher thresholds, true negative rates are significantly higher both given the nature of the selected landscape and that tree cover is more conservatively chosen in the Hansen dataset. A threshold set to 35 achieves a balance of both high true positive and true negative rates with an overall lower level of disagreement between the true images. Therefore, it is at this threshold that Hansen’s dataset best reflects tree cover on the ground.
