Segmentation and classification of high resolution imagery for mapping individual species in a closed canopy, deciduous forest

In this paper we investigate the use of a shadow-based delineation program for identifying segments in imagery of a closed canopy, deciduous forest, in West Virginia, USA, as a way to reduce the noise associated with per-pixel classification in forested environments. Shadows typically cluster along the boundaries of trees and therefore can be used to provide a network of nodes for the delineation of segments. A minimum cost path algorithm, where cost is defined as the cumulative sum of brightness values traversed along the connecting route, was used to connect shadow clumps. To test this approach, a series of classifications was undertaken using a multispectral digital aerial image of a six hectare test site and a minimum cost path segmentation. Three species were mapped: oaks, red maple and yellow poplar. The accuracy of an aspatial maximum likelihood classification (termed PERPIXEL classification) was 68.5%, compared to 74.0% for classification using the mean vector of the segments identified with the minimum cost path algorithm (MEAN_SEG), and 78% when the most common class present in the segment is assigned to the entire segment (POSTCLASS_SEG). By comparison, multispectral classification of the multispectral data using the field-mapped polygons of individual trees as segments, produced an accuracy of 82.3% when the mean vector of the polygon was used for classification (MEAN-TREE), and 85.7% when the most common class was assigned to the entire polygon (POSTCLASS_TREE). A moving window-based post-classification majority filter (POSTCLASS_MAJ5BY5) produced an intermediate accuracy value, 73.8%. The minimum cost path segmentation algorithm was found to correctly delineate approximately 28% of the trees. The remaining trees were either segmented, aggregated, or a combination of both segmented and aggregated. Varying the threshold that was used to discriminate shadows appeared to have little effect on the number of correctly delineated trees, or on the overall accuracy of the multispectral classification, although it did have a notable effect on the proportions of aggregated and segmented trees.