Mapping forest parameters using geostatistics and remote sensing data.

摘要

This work presents a new method for characterising forests with remote sensing data using numerical scene simulations and spatial statistics. The principal study area is Cat Tien National Park, Vietnam. This site has undergone both recent changes in vegetation composition associated with population pressures, as well as historical changes due to military activities during the 1960s and 70s and provides an appropriate location for spatio-temporal monitoring of forest structure. The principal remote sensing data used comprises a set of panchromatic declassified air-photos (1965--1966). The lack of flight details for these makes established techniques for exterior orientation impractical. An alternative means to geo-rectifying these data is therefore presented. This focuses on a new application of a stereomatching algorithm, where a disparity model, related to topographic features, is first built and then co-registered to a geo-referenced elevation model to provide the transformation required to correct the air-photos. These geo-rectified data are then processed for forest parameter extraction. Scene modelling is used to produce simulations of varying ground structure. A geo-optical model is used to capture the shape and size distribution of objects in the scene, and to allow for crown shading on the trees. The scene variogram is considered as a combination of spatial interactions between scene elements (crown and ground), which are described by 'component variograms'. These are examined under differing scene specifications, and used to explore and explain the mechanisms responsible for variations in scene variogram 'range' across multi-spectral data. The scene simulations provide a set of candidate model variograms, derived from physical realisations of scene structure, for use in inverting the experimental scene variogram, where forest structural parameters are derived from the realisation associated with the best fit. Results are presented for the high resolution air-photos, and validated using local image histogram analysis, given the lack of in situ data for the time of acquisition. The method in this context appears to be robust and results suggest it can be applied to large areas for forest parameter mapping. Applications of the method to alternative datasets are also examined, with a consideration of necessary adaptations, given changes in spatial, spectral, angular and temporal sampling. These focus on a set of AirMISR images acquired over the SAFARI 2000 site of Mongu, Zambia. Although these suffer from a poorer spatial resolution, the spectral and angular sampling is greatly improved (4 bands and 9 angles of data). In this case, scene simulations are used in conjunction with measures of local variance (rather than semivariance) for inverting structural parameters. Results presented for this dataset are poor, although this is shown to be attributable to the initial sensor calibration, rather than to the method itself.