Adaptive kernels for texture based analysis of object based classification of forest stands

Conventionally forest stands have been delineated on aerial photographs with the help of human pattern recognition. However, the approach is subjective in nature, and can be costly in terms of time and resources. The results are often inconsistent and inaccurate. Image segmentation has been explored as an alternative solution, providing more objective delineation and extracting more information, increasing the consistency of the results in a reduced time span for lower cost. However, segmentation of an image does not necessarily result in objects which are visually meaningful, or correspond to entities of ecological or management interests. Therefore, it is critical to take into account the accuracy with which image segmentation extracts the shapes and position of real objects. This research used the measures of segmentation “goodness”, to evaluate segmentation procedures and to determine the optimal scales for different segmentation schemes. This study explored an automated method for identifying optimal segmentation scale parameters by ordering and ranking different segmentations schemes. A total of 14 scale parameters (ranging from 10 to 800) were explored for three combinations (segmentation schemes) of remotely sensed data. The results showed that the optimal scale was found to be 60, 80 and 100 for the three segmentation schemes. There is considerable uncertainty attached to the results of automated segmentation processes under conditions of forest heterogeneity. Research has suggested that the image segmentation of multispectral data in combination with LiDAR data can improve the classification of stands on the basis of height and tree species for stand delineation. This study explored object based classification to delineate forest stands using QuickBird image separately, and multispectral image and LiDAR data co-jointly. The results illustrated that a significantly high overall accuracy (82.2 %) was achieved when multispectral image and LiDAR data were co-jointly considered as compare to multispectral image alone (64.3%). Furthermore, the classification scheme based on multispectral image and LiDAR data accurately differentiated both mature and pole stands as compare to classification scheme based multispectral image alone. Traditional classification procedures consider only spectral properties of remotely sensed data. The potential application of spatial properties for automatically classifying remotely sensed data has been less explored. This study also illustrated the use of spatial statistics like Local Moran’s Ii in conjunction with other remotely sensed data in order to improve the classification accuracy of forest stands identification. The optimal segmentation result based on multispectral image, LiDAR data and Local Moran’s Ii was classified. The results showed that there was no significant improvement in overall accuracy. However, in some cases pole stands were classified as mature stands by other segmentation/classification schemes. In contrast, segmentation based on multispectral image, LiDAR data and Local Moran’s Ii resulted in accurate classification by taking into account the texture of stands. The use of an adjustable kernel to estimate the spatial autocorrelation of the remotely sensed data has been given little attention in remote sensing community. However, researchers have suggested that different kernel sizes may be more appropriate for identifying different types of forest stand. This research suggested a novel method of identifying the optimal kernel size for classifying a particular area (pixels/classes) by evaluating data sets of Moran’s Ii layers calculated using different kernel sizes. This study termed this approach as “Moran’s Ii Winner Grid”. Moran’s Ii Winner Grids based on near infra-red band and canopy height model derived from LiDAR were used. This study found that a 3x3 kernel was more effective in small stands with small uniform canopies, in areas with small canopy gaps within and in between mature stands. In contrast, a 7x7 kernel was the winner in areas with larger uniform canopy surfaces especially in areas with young forest stands where these is less or no gap between stands. The results from image classification of the multispectral image and Moran’s I winner grid based on NIR show improvements in the overall classification accuracy.