Douglas J. King

Derivation and validation of Canada-wide coarse-resolution leaf area index maps using high-resolution satellite imagery and ground measurements

Leaf area index (LAI) is one of the surface parameters that has importance in climate, weather, and ecological studies, and has been routinely estimated from remote sensing measurements. Canada-wide LAI maps are now being produced using cloud-free Advanced Very High-Resolution Radiometer (AVHRR) imagery every 10 days at 1-km resolution. The archive of these products began in 1993. LAI maps at the same resolution are also being produced with images from the SPOT VEGETATION sensor. To improve the LAI algorithms and validate these products, a group of Canadian scientists acquired LAI measurements during the summer of 1998 in deciduous, conifer, and mixed forests, and in cropland. Common measurement standards using the commercial Tracing Radiation and Architecture of Canopies (TRAC) and LAI-2000 instruments were followed. Eight Landsat Thematic Mapper (TM) scenes at 30-m resolution were used to locate ground sites and to facilitate spatial scaling to 1-km pixels. In this paper, examples of Canada-wide LAI maps are presented after an assessment of their accuracy using ground measurements and the eight Landsat scenes. Methodologies for scaling from high- to coarse-resolution images that consider surface heterogeneity in terms of mixed cover types are evaluated and discussed. Using Landsat LAI images as the standard, it is shown that the accuracy of LAI values of individual AVHRR and VEGETATION pixels was in the range of 50–75%. Random and bias errors were both considerable. Bias was mostly caused by uncertainties in atmospheric correction of the Landsat images, but surface heterogeneity in terms of mixed cover types were also found to cause bias in AVHRR and SPOT VEGETATION LAI calculations. Random errors come from many sources, but pixels with mixed cover types are the main cause of random errors. As radiative signals from different vegetation types were quite different at the same LAI, accurate information about subpixel mixture of the various cover types is identified as the key to improving the accuracy of LAI estimates. D 2002 Elsevier Science Inc. All rights reserved.

Automated tree crown detection and delineation in high-resolution digital camera imagery of coniferous forest regeneration

Ensuring successful forest regeneration requires an effective monitoring program to collect information regarding the status of young crop trees and nearby competing vegetation. Current field-based assessment methodology provides the needed information, but is costly, and therefore assessment frequency is low. This often allows undesirable forest structures to develop that do not coincide with management objectives. Remote sensing techniques provide a potentially low-cost alternative to field-based assessment, but require the development of methods to easily and accurately extract the required information. Automated tree detection and delineation algorithms may be an effective means to accomplish this task. In this study, a tree detection–delineation algorithm designed specifically for high-resolution digital imagery of 6-year-old trees is presented and rigorously evaluated. The algorithm is based on the analysis of local transects extending outward from a potential tree apex. The crown boundary is estimated using the point of maximum rate of change in the transect data and a rule base is applied to ensure that the point is contextually suitable. This transect approach is implemented in both the tree-detection and crown-delineation phases. The tree-detection algorithm refines the results of an initial local maximum filter by providing an outline for each detected tree and retaining only one local maximum value within this outline. The crown-delineation algorithm is similar to the detection algorithm, but applies a different rule set in creating a more detailed crown outline. Results show that the algorithm’s tree-detection accuracy was better than that using commonly applied fixed-window local maximum filters; it achieved a best result of 91%. For the crown-delineation algorithm, measured diameters from delineated crowns were within 17.9% of field measurements of diameter at the crown base on an individual tree basis and within 3% when averaged for the study. Tests of image pixel spacings from 5 to 30 cm showed that tree-detection accuracy was stable except at the lowest (30-cm) resolution where errors were unacceptable. Delineated crown-diameter accuracy was more sensitive to image resolution, decreasing consistently and nonlinearly with increasing pixel spacing. These results highlight the need for very high resolution imagery in automated object-based analysis of forest regeneration. D 2002 Published by Elsevier Science Inc.

Comparison of pixel-and object-based classification in land cover change mapping

Land use/land cover (LULC) change occurs when humans alter the landscape, and this leads to increasing loss, fragmentation and spatial simplification of habitat. Many fields of study require monitoring of LULC change at a variety of scales. LULC change assessment is dependent upon high-quality input data, most often from remote sensing-derived products such as thematic maps. This research compares pixel- and object-based classifications of Landsat Thematic Mapper (TM) data for mapping and analysis of LULC change in the mixed land use region of eastern Ontario for the period 1995–2005. For single date thematic maps of 10 LULC classes, quantitative and visual analyses showed no significant accuracy difference between the two methods. The object-based method produced thematic maps with more uniform and meaningful LULC objects, but it suffered from absorption of small rare classes into larger objects and the incapability of spatial parameters (e.g. object shape) to contribute to class discrimination. Despite the similar map accuracies produced by the two methods, temporal change maps produced using post-classification comparison (PCC) and analysed using intensive visual analysis of errors of omission and commission revealed that the object-based maps depicted change more accurately than maximum likelihood classification (MLC)-derived change maps.

Airborne digital camera image semivariance for evaluation of forest structural damage at an acid mine site

Abstract A forest downstream of a heavy metal acid tailings area at the KamKotia mine site near Timmins, Ontario shows visible signs of damage which include varied leaf size, leaf discoloration, dead branches, and increased individual tree crown and forest canopy openness. High resolution remote sensing has potential for providing means to spatially and temporally evaluate such damage. In particular, image texture can be used in modeling forest and individual tree structural variations which may result from stress. In this study, an airborne multispectral digital camera system was used to acquire imagery with ground pixel spacing of 0.25 m, 0.5 m, and 1.0 m. Relations of image semivariance measures with field forest structure and health measures were determined. Semivariograms were derived using two sampling techniques: transects in two perpendicular directions, and omnidirectional sampling within pixel matrices. Sampling was conducted over the forest canopy as well as within individual tree crowns. The principal objective of the study was to determine the types of forest canopy and individual tree crown structure and health information captured by semivariograms at the three spatial resolutions. In canopy scale sampling, the 1 m pixel semivariograms were best related to forest canopy closure, stem density, and a visually derived tree stress index. The 0.5 m pixel semivariograms related better to tree crown size and tree height. The transect technique was more sensitive to tree height and the matrix technique produced stronger relations with tree stress. In individual tree crown samples, the 0.25 m pixel semivariograms were well related to tree crown closure. This suggests that semivariograms extracted from the 1 m and 0.5 m pixel images are suitable for mapping structural and textural information related to forest damage at the canopy level while the semivariograms extracted from the 0.25 m pixel images depict information at the tree crown level.

Spatial analysis of radiometric fractions from high-resolution multispectral imagery for modelling individual tree crown and forest canopy structure and health

Research was conducted in a forest adjacent to an abandoned acid mine tailings site to assess forest structural health using high spatial and spectral resolution digital camera imagery. Conventional approaches to this problem involve the use image spectral information, basic spectral transformations, or occasionally spatial transformations of image brightness. This research introduces fractional textures and semivariance analysis of image fractions. They were integrated with conventional image measures in stepwise multiple regression modelling of forest structure (canopy and crown closure, stem density, tree height, crown size) and health (a visual stress index). The goal was to conduct a relative comparison of the potential of the various image variable types in modelling of forest structure and health. Analysis was conducted for both canopy (crowns and shadows) and individual tree crown sample data sets extracted from 10 nm bandwidth spectral bands at three resolutions (0.25, 0.5, 1.0 m). Spatial transformations (texture, semivariogram range) of image brightness (DN) and image fractions (IF) were consistently the most significant and first entered variables in the best models of the forest parameters. At the canopy-scale, despite a limited number of available plots (6), stable models were produced that demonstrated the potential for spatially transformed variables. Semivariogram range explained 88% of the total variation of 9 of the 18 models and represented 56% of the variables used in all models while texture variables explained 51% of model variance in 8 of the 18 models and represented 40% of the variables used. At the tree crown scale (n=31), 88% of the total variation of six of eight models was explained by texture variables and 6% by semivariogram variables. DN and IF variables that were not spatially transformed contributed little to the models at both scales. They represented 4% and 6%, respectively, of the variables used in all models. Spatial information in image fractions and image brightness has proven to be more significant than spectral information in these analyses. Of the spatial resolutions evaluated, 0.5 m consistently produced similar or better models than those using the 0.25 or 1.0 m resolutions. These results demonstrate the potential for integration of spatial transforms of image fractions and raw brightness in high-resolution modelling of forest structure and health. D 2002 Elsevier Science Inc. All rights reserved.

Quantification and reduction of bidirectional effects in aerial cir imagery of deciduous forest using two reference land surface types

Bidirectional effects in airborne remote sensing data are governed by land surface reflectance characteristics, atmospheric scattering and the sun‐object‐sensor angular relationship. These factors introduce brightness variations in the data that render quantitative analysis more difficult. On the other hand, bidirectional reflectance distribution functions (BRDF) of land surfaces provide information about the physical characteristics of the surface. This study has two objectives. The first is to reduce brightness variations in aerial false colour imagery using empirical methods. The second is to test if bidirectional effects must be modelled using sample data only for the land surface type under analysis or can another surrogate land cover type, which may be easier to sample in the imagery, be used. The method is based on multiple view angle imaging of two reference land surfaces using highly overlapping stereo frame format imagery. Reduction of BRDF effects includes derivation of a quantification factor used in a formula based on the Rayleigh scattering function. The performance of the method is determined by evaluating the residual variations in mean digital number (DN) of several test plots located in two overlapping images. The study area is a temperate deciduous forest damaged to varying degrees by a severe ice storm in 1998. For sixteen test plots throughout the forest, it was found that the mean DN values of the plots in the two images became more similar when the correction model was derived from samples of the same deciduous forests. However, a BRDF model derived from samples of vegetated fields nearby reduced the variations between images for some of the sixteen forest plots more than the forest BRDF model. These plots were less damaged, had high canopy cover, and were located on steep slopes oriented towards the sun. They had scattering characteristics more similar to the fields than to the deciduous forest as a whole. Two principal conclusions were derived. First, it is essential that the model used to derive the BRDF correction be from the same land cover as the land cover type under study. Second, the suitability of the quantification factors for the BRDF correction provided additional information about the forest structure and damage level.

Shadow brightness and shadow fraction relations with effective leaf area index: importance of canopy closure and view angle in mixedwood boreal forest

This research examines the impacts of varying canopy closure and view angle on relations of high-resolution digital camera shadow fraction and shadow brightness with mixedwood boreal forest effective leaf area index (LAIe). Results from linear regression analyses revealed weak and insignificant positive relations between shadow fraction and LAIe. Considerable scatter was observed in the relationship and was attributed to differences in canopy closure among plots, as only forest with greater than 80% closure showed a positive relationship between shadow fraction and LAIe. These results emphasized that gap size and frequency were important factors in determining the shadow fraction. Relations between shadow brightness and LAIe were significantly improved over those with the shadow fraction because of decreased data point scatter when closure was less than 80%. Shadow brightness was not adversely affected by canopy closure, while remaining sensitive to species composition. The association of shadow fraction with LAIe was understood to exist through links to the projected shadow area of tree crowns; however, this relation became unstable in more open canopies. With shadow brightness, surrogate information on LAIe was implicitly linked to differences in the transmission of light through deciduous and coniferous tree crowns. An evaluation of view angle geometry effects suggested that bidirectional reflectance impacts on the shadow fraction ‐ LAIe relations were strongest in the forward-scattering direction, but had less effect on regression analysis in the backscattering direction and at nadir.

Riparian marshland composition and biomass mapping using Ikonos imagery

The Ontario Wetland Evaluation System (OWES) employs mostly field-based visual assessments of wetland extent, composition, and productivity as primary indicators in determining which wetlands should be considered provincially significant and protected. A given wetland is generally assigned single attributes that are assumed to represent the whole wetland extent. High spatial resolution satellite remote sensing offers potential to map and monitor spatial variability of given attributes within a wetland. In this study, Ikonos imagery was used to map vegetation composition and biomass in three riparian marshes near Ottawa, Ontario. For vegetation composition mapping, separability and correlation analyses aided the selection of an optimum set of spectral and texture input variables. Several maximum likelihood classification tests for sets of terrestrial and aquatic vegetation classes gave best accuracies from 61% for seven classes to 88% for five classes. As an alternative, neural network classification was tested using various configurations of input variables and data. However, the best results did not match those from the maximum likelihood classification. For biomass mapping, dried green and senescent biomass collected at 75 locations were modelled using stepwise forward multiple regression. The best model produced was the logarithm of green biomass against a combination of texture and spectral variables (R2 = 0.61). It was applied to the image data to map green biomass with an absolute error of 213 g/m2, or approximately 40% of the mean field-measured biomass. Based on this error magnitude, the output map was aggregated into three classes of biomass (high, medium, and low) that showed a strong visual correspondence with the spatial distributions observed in the field. These results indicate strong potential for monitoring of vegetation composition and biomass changes within wetlands which may contribute to improvement of wetlands evaluation and monitoring in Ontario.

Lacunarity analysis to determine optimum extents for sample‐based spatial information extraction from high‐resolution forest imagery

Lacunarity analysis was evaluated as a means to determine multiple pattern scales that are inherent in high‐resolution imagery of forests and to specify an optimal spatial extent for spatial image information extraction. A series of 0.5 m pixel images of temperate hardwood and mixed boreal forests were analysed using lacunarity distributions calculated for spatial extents ranging from 7 m to 40 m. The optimal extent was taken as that which displayed the greatest number of distinct pattern scales. For the temperate hardwood forest dataset, 12–14 m extents were found to be optimal, detecting three pattern scales. For the boreal forest dataset, optimal extents were 14–18 m for five of six plots, detecting two or three pattern scales in each plot. The detected pattern scales ranged from 8 m to 14 m and showed some correspondence to tree crown size, but also responded to clusters of understorey and overstorey trees or to partially exposed tree crowns. The method can aid in determination of the sample extent that best captures the pattern scales present in the imagery. More generally, it can be useful in exploratory analysis of any spatial data for which the fundamental patterns are not known.

Temporal analysis of forest structural condition at an acid mine site using multispectral digital camera imagery

A large abandoned tailings deposit at a mine site near Timmins, Ontario, Canada has produced significant damage in an adjacent forest due to contamination and wind stress. Significant forest structure changes were measured between 1997 and 1999. A multivariate image-based forest structure index (FSI) was developed using canonical correlation analysis of 1997 field and airborne digital camera data. FSI included decreasing canopy closure and leaf area index, and increasing blown down and standing dead structure measures associated with image spectral, textural and radiometric fraction variables. An image model predicting FSI achieved an R 2=0.66. The model equation was then applied to 1999 airborne imagery to predict FSI for each plot. Comparing the 1999 image predicted FSI to that calculated from field data showed that the model was strong in predicting positive or no forest structure changes, but not increased structure degradation. The latter was due to the presence of herbaceous and shrub vegetation that had developed during the two-year period in open plots near the tailings where blow down was significant. The next research phase will derive means to separate these two signals in forests of open overstory.