DIARIO DE CAMPO
4. ANÀLISIS DE LOS RESULTADOS Y DISCUSIÒN
4.2 ANALISIS DE OBSERVACIONES REGISTRADAS EN EL DIARIO DE CAMPO
Remotely sensed imagery records the spectral reflectance from the surface of the Earth, and is a useful source of information about the spatial distribution of vegetation. Applications for mapping the location of vegetation, its quality, quantity and dynamics are numerous and continually expanding. There is also a correspondingly intensive research effort for improving analytical methods (Dungan, 2001). The scale of the land surface unit under observation (pixel size) and spectral range combine to determine the amount of information that can be derived from these sensors; thus an understanding of potential scale effects is critical to effective use of remotely sensed data (Hay et al., 2005).
Remote sensing use within forest inventory has been steadily increasing over the last 60 years. The first use began with black and white aerial photography after the First World War, with major utilisation and colour photography occurring after World War II. There has been an
ever expanding use of satellite remote sensing from the 1970’s, with the spatial resolution, extent, diversity (multi-spectral, hyper-spectral, microwave/radar, laser), and reliability of remote sensing technologies improving rapidly over the last decade (Wulder, 1998). Additionally, the increasing adoption of a range of remote sensing instruments in multi-resource inventories has produced more accurate information, particularly for defining forest boundaries and producing national level maps (Lund, 1998; McRoberts and Tomppo, 2007).
Methods for processing moderate spatial resolution (~20 m+) remote sensing to produce forest and land cover information are well developed and accepted (e.g., Richards et al., 2000; Cihlar, 2000; Donoghue, 2000; Patenaude et al., 2005). Broad scale forest mapping and monitoring worldwide has been undertaken primarily using NASA’s Landsat series, the Advanced Very High Resolution Radiometer (AVHRR) (e.g., Lu et al., 2003), the Satellite Pour l'Observation de la Terre (SPOT) series, and more recently with the Moderate Resolution Imaging Spectroradiometer (MODIS) sensors (Wulder, 1998; Hill et al., 2006). Processing methods for determining forest cover from moderate spatial resolution sensors like Landsat TM, range from band ratio combinations such as NDVI (e.g., Miura et al., 2006) to more complex modelling using, for example, regression based modelling (Lucas, et al., 2006c) or Geo-Optical (GO) and radiative transfer methods (Li and Strahler 1985; Li et al., 1995, McCloy, 2006). With GO methods for example, pixel radiance is modelled as the area-weighted combination of the range of sunlit and shaded tree objects and background components visible to the sensor. GO models are used to estimate the bidirectional reflectance distribution function using discrete 3D objects, where the shape, density and patterns control the reflectance response to illumination and different view angles. In the Li-Strahler GO model a spherical shape is assumed for the partially illuminated tree crowns that make up the vegetation canopy (Jupp and Walker, 1996; Scarth and Phinn, 2000).
The Landsat series of satellites have been in operation since 1972, and have been utilised for a broad range of applications in Australia. The latest sensors - Landsat 5 Thematic Mapper (TM) and Landsat 7 Enhanced Thematic Mapper (ETM) - have 30 m spatial resolution across 7 spectral bands, with Landsat 7 having an additional 15 m panchromatic band (NASA,
2008). A fault developed in Landsat 7 during 2003 resulting in data gaps, rendering its data unusable for National Forest Inventory applications; however more recent research into increasingly sophisticated data blending algorithms incorporating MODIS and Landsat, the SLC-off limitation is becoming less of an issue (Gao et al., 2006; Wulder et al., 2008). This also highlights the requirement for CFMF data collection methods to be technology independent (Principle 4), or at least not substantially reliant on any one sensor (Wood, et al., 2006). In Australia, two major users of Landsat TM imagery for state and national level forest measurement and monitoring are the Queensland Statewide Land-cover And Trees Study (SLATS) program and the Australian Greenhouse Office National Carbon Accounting System (NCAS).
SLATS use Landsat TM imagery from 1989-to-present for estimating the extent and change in woody vegetation. A multiple regression vegetation index developed using field site data sampled throughout Queensland, is used to calculate a gradient of woody foliage projective cover, and seeks to detect woody vegetation to the lowest possible detection limit (QDNRM, 2003; Lucas et al., 2006b). This vegetation index compensates for the difference in background soil colour which can otherwise cause significant overestimation (for black soil) or underestimation (for red soil) of woody vegetation cover. Dry season imagery (July – September) is used to minimise the variability in image quality due to atmospheric haze and cloud, and to provide the maximum differentiation between pasture and forest canopy (Lucas, et al., 2006b). For purposes of vegetation management, SLATS detects all woody vegetation, with a minimum threshold of approximately 7 % foliage projective cover in most cases, but the minimum threshold may be up to 12 % where image quality is poorer (AGO, 2003; QDNRM, 2003). Recently airborne LiDAR data were collected at a number of test sites, and research is underway to use the LiDAR to improve the Landsat TM derived foliage cover estimation models (J. Armston, pers. com.).
NCAS use Landsat TM data to identify change in forest cover status to feed into national carbon accounting models. The primary mission of NCAS is to support Australia’s position in the international development of policy and guidelines for terrestrial greenhouse gas
sinks and emissions. NCAS estimates continental biomass stock and flux at sub-hectare resolution using Geographic Information Systems (GIS) based process models (3PGs) combined with empirical data (Richards and Brack, 2004; Richards and Evans, 2004). Maps of biomass potential are interpolated from mapped multi-temporal productivity layers, calibrated with measurements of mature forest biomass. Biomass accumulation is estimated with simple growth formulae, linked to site productivity (predicted using soil and climate factors) and combined with stand age estimated from disturbance analyses derived from Landsat MSS and TM images spanning 1972-2002 (Richards, 2002). NCAS was developed using a process of continuous improvement, and this allows for enhanced capability for monitoring, scenario modelling, and capability to support climate change mitigation initiatives (Brack et al., 2006). For example, vegetation layers used in the modelling for stratification and allocation of mature biomass estimation functions were derived initially from a combination of the Resource Assessment Commission survey (RAC, 1992) and Carnahan (AUSLIG, 1990) mapping (Richards, 2002). A more recent update includes the National Vegetation Information System data (ESCAVI, 2003).
The NCAS, according to international guidelines, measures a claimed constant minimum threshold of 20 % interpreted tree crown cover at 2 m height, and identifies only verifiable, deliberate land use change (i.e. from a forest to a non-forest use). It was noted in the NCAS documentation that the 20 % crown cover is at the lower limit of cover that can be accurately resolved using Landsat TM data (Richards, 2002). However it should be noted that non-stereo optical data is inherently asymptotic to vertically distributed structural elements (McCloy, 2006), so while the NCAS may claim to utilise a 2 m height threshold, in practise this can only be assumed or roughly estimated based on temporal analyses. Actual fine scale active data sources such as LiDAR would be required to quantitatively apply a specific height threshold to the modelling. Whilst seeking to develop a consistent and comprehensive assessment of vegetation across the continent, these data still contain knowledge gaps and issues with mapping scale, currency, and use (e.g., determining a 2 m height threshold using passive
optical data), which can introduce error when assessed, utilised, or validated at the sub-hectare level.
Synthetic Aperture Radar (SAR) is another potential source of forest information. Knowledge of the information content of this data source acquired over forest environments, and in particular microwave interaction with different components (leaves, trunks, branches) is required to support the retrieval of their biomass, structure and floristic composition at an operational level (Hyyppä et al., 2000). Such knowledge is increasingly important given the deployment of lower frequency spaceborne radar sensors (e.g., Japan’s Advanced Land Observing Satellite (ALOS) Phased Array L-band SAR) to complement the current suite of C- band sensors observing the Earth. Active microwave sensors can penetrate the canopy and so provide information about the entire vertical depth of the forest, as well as being sensitive to a range of forest parameters, including the geometric structure of tree components (Liang et al., 2005).
Another benefit of SAR systems is the ability to make all-weather (although results can be negatively impacted by wet conditions e.g., de Jong et al., 2002) and night-time observations at high spatial resolution, and at a range of frequencies and polarizations. In gaining such knowledge, a number of approaches can be adopted, including the use of empirical relationships (Dobson et al., 1995; Lucas et al., 2000; Le Toan et al., 2004) and modelling by distributing tree components of varying size, geometry and dielectric properties within two-dimensional (2D) layers (e.g., Durden et al., 1989; Lucas et al., 2004; Liang et al., 2005). More recent modelling have utilised 3D cubes (e.g., Sun et al., 2002) and simulated microwave interaction, overall backscatter coefficient, and the magnitude of contributory mechanisms (e.g., single bounce, volume scattering). In the latter case, the cubes or “voxels” (volumetric pixels) have typically been constructed around artificial trees or those that have been measured from field data. However, at the present time these systems and methodological approaches tend to be best suited for structurally homogeneous forest types, and those in the lower biomass range (30-200 Mg ha-1) due to issues of sensor saturation (Imhoff, 1992; Lucas et al., 2006a).
LiDAR remote sensing directly measures both horizontal and vertical spatial elements of forest structure. Many studies utilising either both small footprint (< 0.5 m radius) or large footprint (10 m +) waveform digitization airborne LiDAR, have demonstrated an ability to recover structural elements such as tree and canopy height, canopy cover and volume, canopy height profiles, biomass and basal area at accuracies near equivalent to (and sometimes better than) field survey (Magnussen and Boudewyn, 1998; Lim et al., 2003b; Riano et al., 2004; Gobakken and Naesset, 2005; Lefsky et al., 2005b). LiDAR remote sensing of forests will be covered in more detail in section 2.3.2.
When mapping forest structure, an optimal integration strategy would include finely detailed measurements that field sampling (or field equivalent remotely sensed data) provides, combined with the broad spatial coverage of moderate spatial resolution remote sensing. Although this level of forest structural information cannot be provided by any current single technology, advancements in SAR and LiDAR have the potential to lead to broad-scale mapping of both horizontal and vertical structure in the near future (Reutebuch et al., 2005). However, intermediate scale mapping of forest structure is possible through statistical analyses such a model based sampling, and/or fusion of information from multiple sensors. This process takes advantage of the highly detailed vertical measurements provided by LiDAR: either detailed full waveform-digitizing (where available e.g., SLICER / LVIS (Harding et al., 2001)) or ICESat (Lefsky et al., 2005a), or less detailed but more available small footprints systems (Lim et al., 2003b). The LiDAR is then combined with the broad-scale mapping capabilities of passive optical sensors, for example Landsat TM (Hudak et al., 2002; Wulder and Seemann, 2003; Donoghue and Watt, 2006), or hyperspectral (Bunting and Lucas, 2006; Koetz et al., 2007; Addink et al., 2007; Lucas et al., 2008), or the coarse sensitivity to horizontal and vertical structure afforded by SAR (Reutebuch et al., 2005; Hyde et al., 2006; Lucas et al., 2006a).