The estimated total forest plantation area in New Zealand is approximately 1.70 million ha. Approximately 70% of the plantations with area over 1000 ha are owned by large-scale owners, whose forests undergo regular monitoring and assessment. The remaining 30% of plantation forests are small-scale and are less likely to have regular area and yield assessments. Knowledge of these small-scale forests, especially those under 100 ha, remains very limited, yet they are expected to comprise over 40% of the total radiata pine (Pinus radiata D. Don) harvest volume by 2020. It is critical to better understand the small-scale forest resource in order to plan effectively for marketing, harvesting, logistics and transport capacity for this future resource. A remote sensing solution to small-scale forest description is necessary because conducting a comprehensive survey and field assessment on those patchy forests is impractical. The objective of this research is to apply multi-sensor remote sensing techniques-LiDAR and RapidEye to derive area, stand age and yield information for small-scale forests in New Zealand.

This research compared a factorial combination of two classification approaches (Nearest Neighbour and Classification and Regression Tree) and two remote sensing datasets (RapidEye and RapidEye plus LiDAR) for their ability to accurately classify land cover, specifically planted forest area. The research further determined the optimal modelling approach for deriving forest stand variables - mean top height, basal area, volume and stand age by comparing the performance of two parametric models (multiple linear regression and seemingly unrelated regression) and two non-parametric models (k-Nearest Neighbour and Random Forest) with RapidEye-derived metrics, LiDAR-derived metrics and a combination of both. The optimal mapping and modelling approaches developed on a training area, was then applied to the entire study area, the Wairarapa Region of New Zealand.

CART using a combination of RapidEye and LiDAR metrics outperformed the other three approaches producing the highest accuracy for mapping forest plantations. This method was further examined by comparing the mapped plantations with manually digitised plantations based on aerial photography. Across all sample grids, the mapping approach overestimated the plantation area by 3%. It was also found that forest patches exceeding 10 ha achieved higher conformance with the digitised areas.

LiDAR-derived metrics were found to be more useful in estimating all four forest stand variables relative to RapidEye metrics; combining LiDAR metrics with RapidEye metrics did not provide significant gains (on average 0.2% reduction in RMSE) in variable prediction. Non-parametric models and parametric models performed similarly, likely due to the narrow range of structural characteristics in the collected field data. Overall, multiple linear regression was deemed to be the best option for estimating forest variables for less well known forests as the approach has provided sound and consistent estimation of stand variables and it is relatively easy to understand and interpret.

The optimal area mapping and modelling approaches were applied to the Wairarapa region (594 000 ha), resulting in area and yield description for the region. Overall the mapped plantation area was 3.4% lower than the National Exotic Forest Description (NEFD) recorded plantation area. NEFD is an annual report that provides detailed area and yield description for New Zealand’s plantation forests. The description of the large-scale forests from NEFD is reliable as it is captured directly from surveys collected from forest owners, whereas the description of small-scale forests is less reliable as the information from over half of these forests is imputed indirectly based on nursery studies. Forest stand variables mean top height, basal area, volume and age were modelled for the region using multiple linear regression with LiDAR-derived metrics. Based on the modelled stand variables, the recoverable volume at different ages (yield table) was generated. The yield tables developed using modelled information were within a realistic range and were slightly lower than NEFD yield tables.

Overall, the mapping and modelling approach developed in this research provided a proof of concept for deriving area and yield information using remote sensing data, and is especially relevant for small-scale forests where limited information is currently available. The wood availability from these small-scale forests could be more accurately addressed at a national level using this approach.