Silvilaser 2019 - Poster Presentations »
Mobile Phone Videos for Forest Inventory
At national and provincial levels, information on the state and evolution of forest resources is collected by repeated national forest inventories (NFIs) on well-distributed sample plots of approximately 500 m². Standard quantitative tree and plot-level attributes are the diameter at breast height (DBH, 1.30 m above ground), tree height, and tree density, accompanied by more qualitative information like tree species, soil condition, and tree health. Despite the general progress in remote sensing, NFIs still rely on traditional, time-consuming, manual field measurements using callipers and the like, which limits the number of surveyed plots to typically only 4 or 5 per day and team. How to retrieve these attributes in a cost- and time-efficient way while maintaining the needed accuracy is still an open research question. We suggest to derive them from 3D point clouds reconstructed from mobile phone videos. Capturing image data brings the benefit of a complete documentation of the appearance of the plot during the survey. Usage of videos instead of still images potentially offers two substantial, additional advantages: (1) while image overlap and occlusions must be taken into account for the capture of every still image, video surveys may be as simple as swiftly walking through the plot along a meandering path; (2) since the perspectives of still images typically differ substantially and their orientations are unknown beforehand, point cloud reconstruction is computationally intensive. By contrast, subsequent video frames have perspectives so similar that image features found in one frame can be tracked efficiently across multiple frames. Finally, using mobile phones for video capture has the added benefit of using an ubiquitous sensor. Our video processing provides approximate results. To achieve highest accuracies, we use them as input for a bundle block adjustment that for performance reasons comprises only an automatically selected subset of video frames. This yields a coloured point cloud of the plot from which we automatically derive tree counts, tree positions, and DBHs. Two forest plots in Austria serve as test sites that cover different forest environments. It turns out that capturing a suitable mobile phone video of a forest plot can be done very quickly (approximately 15 minutes per plot), and first results seem very promising, which we will compare to those obtained from TLS measurements and from classical, manual field measurements.