Creating a virtual tropical forest from three-dimensional aerial imagery to estimate carbon stocks

Given the interest in implementing land-use change and forestry projects for mitigating carbon dioxide emissions, there is potentially a large demand for a system to measure carbon stocks accurately and precisely in a cost-effective manner. As terrestrial ecosystems tend to be heterogeneous, a large number of sample plots could be needed to attain the regulatory-required levels of precision, thus resulting in a costly process. A potential way of reducing costs of measuring the carbon stocks of forests is to collect the key data remotely. We have designed a system (a multispectral three-dimensional aerial digital imagery system, M3DADI) that collects high-resolution overlapping stereo imagery (<= 10 cm pixels) from which we can distinguish individual trees or shrubs. In essence, we created a virtual forest that we used to measure crown area and heights of all plant groups. We used this M3DADI system to estimate the carbon stocks in aboveground biomass for the pine savanna in the Rio Bravo Carbon Sequestration Pilot Project in Belize. Seventy-seven plots were established on the images, and using a series of nested plots we digitized the crown area and heights of pine and broadleaf trees, palmettos, and shrubs. Based on standard destructive harvest techniques, we obtained highly significant allometric regression equations between biomass carbon per individual and crown area and height. Combining the image-plot data with the allometric equations resulted in a mean carbon stock of 13.1 Mg/ha with a 95% confidence interval of 2.2 Mg C/ha or +/- 16% of the mean. The coefficient of variation was high for all vegetation types (range of 31-303%), reflecting the highly heterogeneous nature of the system. We estimated that 202 plots would need to be installed to achieve a 95% confidence interval of +/- 10% of the mean. We compared the cost-effectiveness of the M3DADI approach with conventional field methods based on the total person-hours needed by both approaches to collect the same set of data for 202 plots. We found that the conventional field approach took about three times more person-hours than the M3DADI approach.