As remote sensing data continues to proliferate, development and assessment of methods that generate predictions of forest attributes is needed to inform operational use. For several decades, lidar data collected from aerial platforms has informed assessments of forest resources, but many opportunities remain for the technology. We examine and develop methodology for three different problems facing forest management inventories.
In our first manuscript, we assess three different classification methods for identifying stands for commercial thinning operations in southwestern Oregon, USA using a set of fixed radius plots and a coincident aerial lidar acquisition. We also assess the impact of sample size using a simulation procedure. We found that random forests and a newly developed gradient boosting algorithm exhibited moderate performance for commercial thinning classification. We also observed that performance of all classification methods stabilized at sample sizes between 200 and 300, which may be an attainable sample size in some operational forest inventories. This study was motivated by the lack of literature that examined prediction of categorical variables such as management classes, rather than prediction of continuous forest structural variables, as a potential method for assisting forest management planning decisions. We anticipate a number of extensions to other management-oriented classifiers, such as pre-commercial thinning and various other silvicultural objectives, as future studies.
In the second manuscript we compared a segmentation-based method against an area-based approach method for generating stand-level predictions of forest attributes. Particularly, we leverage small area estimation methods to produce model-based mean squared errors for stand-level predictions. The analysis suggests that the segmentation-based method tends to produce lower mean squared errors in stands where sample sizes increased due to tree segmentation. Furthermore, models based on detected segments were less prone to extrapolation than models produced using the area-based method. However, area-based models generally produced lower mean squared errors for stands that did not contain sampled population units. This study was motivated by a lack of investigation into the use of tree-segmentation methods for producing stand-level predictions of forest attributes, which is a typical objective of many forest management inventories. We believe that this manuscript lays the foundation for continued assessment of alternative tree-segmentation methods with rigorous assessments of prediction error.
The final manuscript employed the multivariate Fay-Herriot model, a recent theoretical development in the small area estimation literature, for producing stand-level predictions of forest attributes. We compared bivariate Fay-Herriot models to their univariate counterparts for five forest attributes and observed that, for at least one bivariate pairing, stand-level mean squared errors were reduced for both sampled and unsampled stands. Additionally, we identify the uniformity and strength of correlation among stand-level direct-estimators as an important indicator of the success of a bivariate model over a univariate model. We plan to conduct a future study that leverages the multivariate Fay-Herriot model for use in a time-series analysis to unify remote sensing and field data collected at disparate times.