Mapping forest vegetation for the western United States using modified random forests imputation ofFIAforest plots

Abstract

Abstract. Maps of the number, size, and species of trees in forests across the western United States are desirable for many applications such as estimating terrestrial carbon resources, predicting tree mortality following wildfires, and for forest inventory. However, detailed mapping of trees for large areas is not feasible with current technologies, but statistical methods for matching the forest plot data with biophysical characteristics of the landscape offer a practical means to populate landscapes with a limited set of forest plot inventory data. We used a modified random forests approach with Landscape Fire and Resource Management Planning Tools (LANDFIRE) vegetation and biophysical predictors to impute plot data collected by the US Forest Service's Forest Inventory Analysis (FIA) to the landscape at 30-m grid resolution. This method imputes the plot with the best statistical match, according to a "forest" of decision trees, to each pixel of gridded landscape data. In this work, we used the LANDFIRE data set for gridded input because it is publicly available, offers seamless coverage of variables needed for fire models, and is consistent with other data sets, including burn probabilities and flame length probabilities generated for the continental United States. The main output of this project is a map of imputed plot identifiers at 30 × 30 m spatial resolution for the western United States that can be linked to the FIA databases to produce tree-level maps or to map other plot attributes. In addition, we used the imputed inventory data to generate maps of forest cover, forest height, and vegetation group at 30 × 30 m resolution for all forested pixels in the western United States, as a means of assessing the accuracy of our methodology. The results showed good correspondence between the target LANDFIRE data and the imputed plot data, with an overall within-class agreement of 79% for forest cover, 96% for forest height, and 92% for vegetation group.

Innovation in rangeland monitoring: annual, 30 m, plant functional type percent cover maps for U.S. rangelands, 1984–2017

“…), is robust to overfitting (each regression tree retains an independent fraction of the data [0.368] for validation, known as out‐of‐bag [OOB] samples) and has grown in prominence for mapping land cover with a multitude of small‐scale and regional studies and applications (Riley et al. , Azzari and Lobell , West et al. , Anderson et al.…”

“…Random Forests (Breiman 2001) is a non-parametric machine learning method that utilizes an ensemble of regression trees, has been shown to have higher classification accuracy than simple regression methods (Belgiu andDr agut ß 2016, G omez et al 2016), is robust to overfitting (each regression tree retains an independent fraction of the data [0.368] for validation, known as out-ofbag [OOB] samples) and has grown in prominence for mapping land cover with a multitude of small-scale and regional studies and applications (Riley et al 2016, Azzari and Lobell 2017, West et al 2017, Anderson et al 2018. Random Forests has the ability to model complex non-linear interactions across predictors, leveraging the large quantity of high spatial resolution plot cover estimates with the vast suite of spatiotemporal and static predictor variables.…”

Modeling Fuel Treatment Leverage: Encounter Rates, Risk Reduction, and Suppression Cost Impacts

“…To do this, we relied on previous research mapping treatment opportunities by [33], specifically using their Scenario D, which offered the loosest constraints on treatments (allows access to all timber within 610 m of existing roads on slopes < 35%, and all timber within 305 m of existing roads on slopes < 50%). We then used random forest modeling to assign each pixel on the landscape to a unique tree list that corresponds to an existing Forest Inventory and Analysis (FIA) plot following the methodology of [63]. FIA measures the size and species of each tree on each plot, which provides the basis for these tree lists.…”

Cumulative effects of wildfires on forest dynamics in the eastern Cascade Mountains, USA

“…Although field-based plot inventories provide fine-scale resolution on ecological change at stand scales and the cumulative effects of these on broad regional scale trends (Reilly and Spies 2015), they are generally limited to a single relatively short time period, lacking the sampling intensity to characterize dynamics spatially and at landscape scales. Remote-sensing and predictive vegetation mapping techniques can link dynamics across spatial and temporal scales (e.g., gradient nearest neighbor [GNN; Ohmann and Gregory 2002], k-nearest neighbor [k-NN;Chirici et al 2016], Random Forest [Riley et al 2016]), and are essential for assessing regional scale dynamics and trends in biodiversity (Staver et al 2011, Ohmann et al 2012, Davis et al 2015, Haugo et al 2015. These techniques have their own limitations and sources of uncertainty that are frequently overlooked in ecological studies, but new methods that explicitly incorporate uncertainty and improve confidence in remotely sensed estimates of land use change and forest dynamics are available (Olofsson et al 2013).…”