Development of regional models of Pinus radiata height from GIS spatial data supported with supplementary satellite imagery

Managers of forest resources require detailed information about their crops for planning, management and inventory purposes. Traditionally, forest inventories have been carried out through collection of ground-based measurements describing relevant crop features such as stand density, height and volume. However, as these measurements are usually taken over a limited number of plots, and then extrapolated to entire stands, they do not account for spatial variation across stands. Use of remotely sensed data within forest inventories overcomes this limitation.

Airborne light detection and ranging (LiDAR) has proved to be very useful at characterising the forest canopy in three dimensions. Since the first application of LiDAR in forestry over a decade ago (Nilsson 1996), the technology has been widely used to spatially quantify variation in tree height and crown dimensions at resolutions ranging from the stand level (Hall et al. 2005; Naesset and Bjerknes 2001), plot level (Holmgren et al. 2003; Hyyppä et al. 2001; Lim and Treitz 2004; Popescu et al. 2004) to individual tree level (Chen et al. 2006; Coops et al. 2004; Holmgren and Persson 2004; Persson et al. 2002; Popescu and Zhao 2008; Roberts et al. 2005; Yu et al. 2004).

The accuracy of LiDAR-derived estimates is reported to be similar to or better than manual field measurement methods for estimating tree height and volume (Naesset 2002; Watt 2005). Data obtained using LiDAR are now used operationally in Finland, Norway, Sweden, Spain, Chile, and Australia to provide forest estimates at the compartment level (Eid et al. 2004). Due to the cost of flying, application of LiDAR, is usually limited to local scales. While regional scale data is sometimes utilised (Zhao and Popescu 2009), repeat annual acquisition at this scale is generally cost prohibitive (Hudak et al. 2002).

Although not as accurate as LiDAR the use of satellite remote sensing does provide a far more cost effective means of predicting crop height. Numerous studies have shown that multispectral satellite data can be utilised in models to map important forest attributes over the land surface (McRoberts and Tomppo 2007; Donoghue and Watt 2006; Shamsoddini et al. 2013; Ozdemir 2008; Kayitakire et al. 2006). However, there are limitations to using remotely sensed data to predict stand dimensions across the entire rotation. Many studies have shown that relationships with biomass and leaf area index tend to saturate after canopy closure when spectra or indices are utilised as sole predictors of stand dimensions (Donoghue and Watt 2006; le Maire et al. 2011). Indices are often favoured as they are likely to be less sensitive to changes in solar illumination and atmospheric conditions than individual bands. Although it is generally accepted that use of remotely sensed data relies on local calibration, little research has evaluated the robustness of relationships when they are applied across different locations.

Spatial information held in geographic information systems (GIS) is also useful for developing models of tree growth. This approach has considerable merit as forest managers generally have access to GIS and a number of potentially useful predictive variables. Rapid increases in the capability of GIS over recent years have seen the development of spatial surfaces covering a diverse range of climatic (Leathwick and Stephens 1998; Leathwick et al. 2003; Tait et al. 2006) and edaphic variables (Watt and Palmer 2012) that in combination with stand information such as age, can be used to model tree productivity. These environmental surfaces have been successfully used to improve empirical growth models (Dzierzon and Mason 2006) and to develop spatial surfaces describing indices of plantation productivity, at a given age, such as Site Index (Palmer, Hock et al. 2009; Watt et al. 2009; Palmer et al. 2012).

Prediction of forest dimensions using data obtained from GIS surfaces and satellite imagery has been less widely researched. Previous research has shown gains in predictive power when models of important stand dimensions are created using spectral data and stand age (le Maire et al. 2011; Gebreslasie et al. 2010; Zheng et al. 2004), relevant indices of productivity (Gebreslasie et al. 2010) and bioclimatic data (Baccini et al. 2008). The present authors are unaware of any research that has investigated the relative contributions of satellite imagery and all three described components of GIS spatial data in prediction of forest dimensions. The objective of this research was to develop models of Pinus radiata height using GIS spatial data supplemented with high resolution satellite imagery.

This study clearly showed that precise models of tree height can be developed using stand age and a GIS surface describing Site Index. Supplemental information describing REVI, obtained from fine resolution satellite imagery, significantly improved the models but this information was of secondary importance to use of GIS layers. Although of secondary importance, use of REVI does provide a temporally updateable component to the height models that could potentially account for the impact of abiotic and biotic factors on tree height.

At both study sites, age was the strongest determinant of H_m. Comparison of partial response functions show that age exceeded the contribution of other variables by at least nine-fold at Kaingaroa and six-fold at Tairua. The strength of the age–height relationship was largely attributable to the relatively narrow range in Site Index at both sites, which varied by ca. 7 m. Site Index often varies more widely at a forest level, and even more widely at the national level, with an almost three-fold range (from 15.1 to 42.5 m). The relationship between LiDAR height and age will be weaker in forests with greater variability in Site Index.

Inclusion of Site Index in the model allows for more generalised prediction. Empirical height growth models typically use Site Index to scale age-height relationships between sites with differing productivity. Consequently, from a theoretical point of view empirical selection of this variable in the model was sound and reinforced the accuracy of the Site Index layer that was primarily based on productivity relationships with air temperature and rainfall (Palmer, Watt et al. 2009). Age and Site Index should be included within height productivity models given the critical importance of these two variables. In situations where Site Index is not available, use of key environmental variables (such as air temperature and mean annual rainfall) may provide useful surrogates for Site Index.

After accounting for age and Site Index, the strongest correlation was between REVI and H_m. The REVI uses a narrow spectral band width along the red edge band that is sensitive to changes in stress and chlorophyll concentrations of different types of vegetation (Chappelle et al. 1992; Gitelson et al. 1996; Carter and Knapp 2001). Previous research using an aerial platform has also identified the red-edge band as providing useful information on photosynthetic vigour (Curran et al. 1990).

Gains in precision were greatest when REVI was combined with both age and Site Index. Within the fitting dataset, relatively substantial gains were obtained over the age-only model at both Kaingaroa (RMSE = 1.29 vs. 1.73 m) and to a lesser extent at the Tairua site (RMSE = 2.96 vs. 3.13 m). The validation process confirmed the importance of REVI at the Kaingaroa site. However gains obtained through adding REVI and Site Index to the base model with age did not result in model improvement at the Tairua site. This is possibly attributable to increased presence of topographic shadow due to the lower sun angle (35°) and undulating terrain over the Tairua site. These factors effectively mitigate the shadow correction process in areas of high relief. This is evident in site differences in REVI where REVI was markedly lower over the Tairua site.

The functional form of the models was similar for both study sites, despite site differences in the magnitude of REVI. This result provides an indication that the same factors influence the relationship at each site. Previous research has shown that REVI is prone to saturation once the canopy has closed (Donoghue and Watt 2006; le Maire et al. 2011). This study demonstrates utility of REVI at ages beyond canopy closure if the variable is combined with age and other spatially derived layers that describe environmental processes and interactions.

Inclusion of satellite-based metrics, such as REVI, in height models allows for fine-scale spatial predictions that can be temporally updated. For example, within stand variation in productivity is not accounted for by age as this is a compartment-level variable. Similarly, while Site Index partially accounts for productivity variation within stands (at a scale of 100 m²) this variable is a static measure that is not sensitive to inter- annual variation in site conditions. In contrast, REVI is available at spatial resolutions that allow for fine-scale predictions of productivity. As the REVI layer can be continuously updated, predictions from repeat acquisitions are sensitive to changes in forest structure caused by growth or changes to the forest canopy caused by wind, snow, fire or disease (Watt and Watt 2011).

The use of a difference index or ratios such as REVI are of particular interest as they provide an effective way to correct for variations in the reflectance values that may be caused by differences in atmospheric or sun illumination conditions. This partially negates the need to conduct pre-processing steps such as atmospheric corrections. Findings from the current study show that the functional form between REVI and height was very similar between sites although the satellite information used was acquired at different times. These results reinforce the utility of REVI as a general predictor of height growth using data acquired under different atmospheric conditions.

In contrast to the results presented here, previous research has shown that mean height of P. radiata can be accurately predicted using only high resolution satellite imagery data (Shamsoddini et al. 2013). Shamsoddini et al. (2013) show strong correlations between mean height and reflectance bands (R² of 0.82) and band ratios (R² of 0.78) from a WorldView-2 multispectral image. The most precise model of mean height was developed using textural attributes and had a coefficient of determination of 0.93 (Shamsoddini et al. 2013). The approach used by Shamsoddini et al. (2013) is not reliant on up-to-date GIS records, that include recently harvested areas, as the relationships used for prediction are more sensitive to variation between bare ground and forested areas, than those used here that were primarily reliant on stand age derived from the GIS. The approach used in the present study is useful in well managed plantations where managers have access to spatial surfaces describing variation in site productivity such as Site Index. Although the approach outlined here is reliant on up to date stand records a methodology has been developed using RapidEye imagery to update recently harvested areas and segregate different species within the stand GIS (Watt and Watt 2011; Watt and Watt 2012). This methodology could be used to update the stand GIS immediately before the models described in this paper are applied to estimate height. The use of a single variable such as REVI from satellite imagery is justified as the simplicity of the relationship may allow generalisation that is not possible with more complex models that include a number of independent variables.

Further research should be undertaken across broader spatial scales that include greater gradients in Site Index to further test the generality of the models developed here. Development of generally applicable regional models of height would be of considerable use to the forestry sector as this approach provides a means of more frequently mapping and monitoring of crop performance. Once this benchmark has been established then subsequent acquisition of imagery should be able to identify clearfell areas not updated in the stand GIS or areas with abiotic or biotic damage. This information would be useful to both resource planners and managers. Frequent coverage from satellite-based systems will also assist in the focussing of LiDAR surveys around more heterogeneous areas (Watt and Watt 2012).

In conclusion, the retrieval of forest structure information over the lifetime of a planted forest is best achieved by the integration of information from GIS layers with satellite, airborne and ground-based measurements. Through such integration of data into a GIS, foresters have a powerful tool that is capable of providing a rapid overview of a large forest estate. These data provide both a pictorial view of the forest and quantitative analysis that can underpin better informed forest management decisions.