PRECIPITACION PLUVIAL
ETAPA FISIOLOGICAMAIZ (Local)
5.1.4. El cultivo de Zapallo 1. Días a la Emergencia
Fusion between LiDAR and other data sources is becoming a research topic in itself. Some studies concern the simultaneous use of LiDAR and multi or hyperspectral datasets, while others consider the combination of 3D information obtained from LiDAR and photogrammetric techniques (St-Onge et al., 2008) using both aerial and satellite imagery. The combination of the 3D LiDAR and 2D spectral information is an area of significant potential. For example the expected combination of information from optical and LiDAR data will be better able to delineate tree crowns from above (Holmgren et al., 2008; Leckie et al., 2003a). Another area of research has involved the use of combining optical multispectral, LiDAR and radar data. Hyde et al. (2006) outline an approach combining such information sources statistically in order to extract forest structure information for wildlife habitat analysis. The results concluded that LiDAR was the best single sensor for estimating canopy height and biomass. With the addition of multispectral data, improvements were made in the estimation of tree structure. Hyde et al. (2006) concluded by saying that the structural metrics extracted from LiDAR combined with radar were essentially redundant.
health and spatial geometry of the tree crown is provided by multi or hyperspectral imagery. LiDAR however, provides data concerning tree height, 3D crown shape and texture or outline (Hyyppä et al., 2004; Leckie et al., 2003a). It should be noted that LiDAR and optical data do not have to be collected at the same time (Hyyppä et al., 2004). Non-coincident data capture, for example, optical imagery collected in the vegetation growth season with LiDAR data captured over the vegetation in leaf-off conditions, allow obvious advantages in winter months where lasers can penetrate the canopy of deciduous trees so that vertical structure can be better discerned (Brandtberg et al., 2003).
Only a small amount of literature has been published on the topic of fusing different sensor technologies together for forest applications (Hyyppä et al., 2004). Of the published literature concerning data fusion, examples have been presented where high resolution optical data have proven to better outline tree crowns in dense forest situations when compared with LiDAR. LiDAR data however, can reduce the commission errors that occur in open stands with optical imagery, for example the application of a height filter to remove sub-canopy vegetation (Hyyppä et al., 2004; Leckie et al., 2003).
The following sections identify those fusions of multi or hyperspectral and LiDAR remote sensing case-studies which extract or estimate those forest condition indicator metrics stated in Table 2.4.
2.3.3.1 Forest structure
Lucas et al. (2008b) utilised a combination of CASI-2 multispectral data at a nominal spatial resolution of 1m and covering the spectral range of visible to near-infrared, and DR LiDAR data acquired using an Optech 1020 scanner with a sample density of a point every ~1m. The aim of this research was to develop a method to extract estimates of biomass for individual trees for mixed species forests in Queensland, Australia. Individual trees were identified using a combination of the tree crowns delineated automatically through segmentation using the CASI data and stems located using a LiDAR height scaled canopy openness index (HSCOI). Tree species information was then extracted for each of the identified tree crown objects using the multispectral data, unless it was a suppressed tree. The component for biomass of individual trees was estimated using LiDAR-derived height and stem diameter as input to species-specific allometric equations. These estimates corresponded to plot-based estimates with an R2 value of 0.56. Additionally, a second approach utilised a jack-knife
linear regression using LiDAR-derived heights and crown cover at the plot scale and
produced more robust estimates of biomass (R2=0.90). A number of issues were highlighted
in terms of over/under-estimations from the LiDAR data dependent on forest cover species type and stem density due to the complexity of the woodland and to the sometimes poor correspondence with objects generated through segmentation.
Popescu et al. (2004) proposed a fusion of small-footprint LiDAR and multispectral data to estimate timber volume and biomass at the plot-level in deciduous and coniferous forests in Virginia, USA. Individual tree heights and crown diameters were estimated using ITC algorithms. An assumption was made that there was a relationship between the height of the tree and its crown size. A regression analysis was then performed to relate field measures of tree DBH and height against those retrieved from the remote sensing data. Regression models and cross validation were then used to estimate inventory data including volume (R2
deciduous 0.39; coniferous 0.83) and biomass (R2deciduous 0.32; coniferous 0.82).
A number of studies have focused upon combining LiDAR and spectral data for the purpose of modelling fire behaviour. In order to satisfy the criteria laid out for fire monitoring applications, factors such as canopy height, vegetation type, dead and live fuel load, and percent canopy cover must be estimated. Mutlu et al. (2008) utilised raster products from multispectral Quickbird satellite image bands and multiple LiDAR height bins and canopy height for characterising forest fuels, for a study site in eastern Texas, USA. A combination of Principle Components Analysis and minimum noise fraction was used to remove the least relevant raster products in an attempt to improve supervised classification accuracy. This resulted in an accuracy of 90.1% with the fusion of airborne LiDAR and satellite Quickbird imagery.
2.3.3.2 Forest composition
LiDAR data can be used to predict the species of individual trees, at least in boreal forests with relatively few tree species (Hyyppä et al., 2008). Such predictions can be improved using a fusion of LiDAR and optical imagery. However, dense LiDAR data have been shown to produce accurate tree-level classification, even without optical imagery (Hyyppä et al.,
Hill and Thompson (2005) outline a method utilising both airborne hyperspectral data acquired using the HyMap sensor, at 4m spatial resolution covering the spectral range from visible to shortwave infrared (0.437-2.486μm), and discrete return LiDAR data, from an Optech ALTM 1210, with an average point density of one point for every 4.83m2. Their
research investigates the ability of the combined dataset to generate unique thematic classes based upon the tree and shrub species composition and vegetation structure, for a site in Cambridgeshire, UK. Classification of the ten various thematic classes was based on the National Vegetation Classification (NVC) scheme for woodlands and scrub. This was achieved using principle components analysis and LiDAR-derived canopy height models. Automated segmentation algorithms were then applied to identify spatial groupings with similar characteristics. It should be noted that this method does not provide information on ground flora composition.
Dalponte et al. (2008) proposed an alternative method in fusing high spatial resolution airborne hyperspectral and LiDAR remote sensing for classification of tree species in complex environments, in this case a nature reserve in the Po Plain, Italy. A software system was produced to provide inputs from hyperspectral image bands and LiDAR-derived canopy height and intensity raster layers for two tree species classification approaches, these were Gaussian maximum likelihood and Support Vector Machines. A total of 19 tree species classes were extracted and assessed against field validation data. The combination of the two data sources resulted in increased classification accuracy, over using one data source only, particularly in relation to the discrimination of very similar species. The kappa accuracies obtained with different classifiers were as high as 0.89 when incorporating hyperspectral and LiDAR layers.
Simonson et al. (2013) addressed the potential for remote sensing with regard to Natura 2000 habitat monitoring objectives. The potential benefits include the cost-effective production of habitat distribution mapping; in addition to providing biophysical indicators of functioning relevant to favourable conservation status such as LAI and vegetation fractional cover amongst others. The research tests the complementarity of multispectral and DR LiDAR in providing a robust indicator of conservation status, in this case an estimate of species richness. A raster-based analysis was conducted upon LiDAR data, where raster layers were derived for measures of image texture and the maximum, minimum, standard deviation of the first return heights and mean of last return heights. Dimension reduction was performed upon
the multispectral data in the form of Principle Components Analysis (PCA). A supervised maximum-likelihood classification was then performed upon the combined dataset creating a total of 11 land-cover classes relating to semi-natural forest, plantation forest, agricultural land, rocky surfaces and urban environments. A land cover classification accuracy of 70% was achieved using the combined dataset. Then utilising known relationships (r = 0.75, p = 0.001) between mean vegetative height, overstorey and understorey species richness, and spatial aggregation calculations using FRAGSTATS, were combined to provide a proxy indicator was created for habitat condition. The result of this calculation was a map with three classes relating to high, medium and low condition.
2.3.3.3 Summary of attributes
Example case studies concerning the combination of hyperspectral and LiDAR datasets are limited, especially for the application area of forest research. Of those relevant example case studies mentioned here for forest condition assessment, many were concerned with the enhancement of tree species classification. Other examples included the research of Lucas et al. (2008b) and Popescu et al. (2004) into the extraction of individual tree parameters and estimates of biomass. Of particular note was the research performed by Simonson et al. (2013) in order to provide forest biophysical monitoring information for a conservation initiative.