Abstract
Background: Global canopy height models are becoming prolific yet require evaluation across New Zealand's diverse vegetation types to assess their accuracy and applicability. Accurate measurement of canopy height is crucial for estimating above-ground woody biomass, which is essential for modelling carbon emissions and sequestration in the context of climate change. These models generally rely on remote sensing data and machine learning techniques, with Light Detection and Ranging (LiDAR) technology commonly employed for precise measurement.
Methods: This study validated the three latest global canopy height models, each provided at a different resolution: 30-metre, 10-metre, and 1-metre. We assessed the accuracy of the selected models by comparing them against canopy height estimates derived from local Airborne Laser Scanning (ALS) datasets, which served as our reference data. Eleven regions across New Zealand were selected based on ALS data availability, encompassing five vegetation and land cover types. Our methodology involved utilising and automating the processing of large New Zealand ALS datasets. To align resolutions for comparison, the reference canopy height was calculated by aggregating average or maximum heights at 10 and 30 m spatial resolution. Model performances were assessed using statistical metrics, including root-mean-square error (RMSE), bias, and R².
Results: Overall, all models exhibited relatively low R² values, indicating limited capture of canopy height variability. The Potapov 30-metre model performed best with average aggregation in shorter vegetation. In contrast, the Lang 10-metre model showed improved accuracy with maximum aggregation, particularly in taller vegetation, but visual boundaries between different vegetation types were not as distinct. The Tolan 1-metre model provided a balanced approach, minimising biases in lower heights but underestimating taller canopies. Results highlight model-specific strengths for varying vegetation structures and the sensitivity of performances to aggregation methods applied to high-resolution reference ALS data.
Conclusions: All three global canopy height models exhibit varied performance across New Zealand's vegetation types. The findings highlight the importance of vegetation-specific applications to optimise each global model’s accuracy. Currently, these models are suitable for carbon accounting efforts as supplementary tools rather than replacements for existing methodologies.