John W. Coulston

Modeling Percent Tree Canopy Cover: A Pilot Study

Tree canopy cover is a fundamental component of the landscape, and the amount of cover influences fire behavior, air pollution mitigation, and carbon storage. As such, efforts to empirically model percent tree canopy cover across the United States are a critical area of research. The 2001 national-scale canopy cover modeling and mapping effort was completed in 2006, and here we present results from a pilot study for a 2011 product. We examined the influence of two different modeling techniques (random forests and beta regression), two different Landsat imagery normalization processes, and eight different sampling intensities across five different pilot areas. We found that random forest out-performed beta regression techniques and that there was little difference between models developed based on the two different normalization techniques. Based on these results we present a prototype study design which will test canopy cover modeling approaches across a broader spatial scale.

On-the-Fly Massively Multitemporal Change Detection Using Statistical Quality Control Charts and Landsat Data

One challenge to implementing spectral change detection algorithms using multitemporal Landsat data is that key dates and periods are often missing from the record due to weather disturbances and lapses in continuous coverage. This paper presents a method that utilizes residuals from harmonic regression over years of Landsat data, in conjunction with statistical quality control charts, to signal subtle disturbances in vegetative cover. These charts are able to detect changes from both deforestation and subtler forest degradation and thinning. First, harmonic regression residuals are computed after fitting models to interannual training data. These residual time series are then subjected to Shewhart X-bar control charts and exponentially weighted moving average charts. The Shewhart X-bar charts are also utilized in the algorithm to generate a data-driven cloud filter, effectively removing clouds and cloud shadows on a location-specific basis. Disturbed pixels are indicated when the charts signal a deviation from data-driven control limits. The methods are applied to a collection of loblolly pine ( Pinus taeda) stands in Alabama, USA. The results are compared with stands for which known thinning has occurred at known times. The method yielded an overall accuracy of 85%, with the particular result that it provided afforestation/deforestation maps on a per-image basis, producing new maps with each successive incorporated image. These maps matched very well with observed changes in aerial photography over the test period. Accordingly, the method is highly recommended for on-the-fly change detection, for changes in both land use and land management within a given land use.

Fitting the Multitemporal Curve: A Fourier Series Approach to the Missing Data Problem in Remote Sensing Analysis

With the advent of free Landsat data stretching back decades, there has been a surge of interest in utilizing remotely sensed data in multitemporal analysis for estimation of biophysical parameters. Such analysis is confounded by cloud cover and other image-specific problems, which result in missing data at various aperiodic times of the year. While there is a wealth of information contained in remotely sensed time series, the analysis of such time series is severely limited due to the missing data. This paper illustrates a technique which can greatly expand the possibilities of such analyses, a Fourier regression algorithm, here on time series of normalized difference vegetation indices (NDVIs) for Landsat pixels with 30-m resolution. It compares the results with those using the spatial and temporal adaptive reflectance fusion model (STAR-FM), a popular approach that depends on having MODIS pixels with resolutions of 250 m or coarser. STAR-FM uses changes in the MODIS pixels as a template for predicting changes in the Landsat pixels. Fourier regression had an R2 of at least 90% over three quarters of all pixels, and it had the highest RPredicted2 values (compared to STAR-FM) on two thirds of the pixels. The typical root-mean-square error for Fourier regression fitting was about 0.05 for NDVI, ranging from 0 to 1. This indicates that Fourier regression may be used to interpolate missing data for multitemporal analysis at the Landsat scale, especially for annual or longer studies.

Status and future of the forest health indicators program of the USA

For two decades, the US Department of Agriculture, Forest Service, has been charged with implementing a nationwide field-based forest health monitoring effort. Given its extensive nature, the monitoring program has been gradually implemented across forest health indicators and inventoried states. Currently, the Forest Service’s Forest Inventory and Analysis program has initiated forest health inventories in all states, and most forest health indicators are being documented in terms of sampling protocols, data management structures, and estimation procedures. Field data from most sample years and indicators are available on-line with numerous analytical examples published both internally and externally. This investment in national forest health monitoring has begun to yield dividends by allowing evaluation of state/regional forest health issues (e.g., pollution and invasive pests) and contributing substantially to national/international reporting efforts (e.g., National Report on Sustainability and US EPA Annual Greenhouse Gas Estimates). With the emerging threat of climate change, full national implementation and remeasurement of a forest health inventory should allow for more robust assessment of forest communities that are undergoing unprecedented changes, aiding future land management and policy decisions.

From sink to source: Regional variation in U.S. forest carbon futures

The sequestration of atmospheric carbon (C) in forests has partially offset C emissions in the United States (US) and might reduce overall costs of achieving emission targets, especially while transportation and energy sectors are transitioning to lower-carbon technologies. Using detailed forest inventory data for the conterminous US, we estimate forests’ current net sequestration of atmospheric C to be 173 Tg yr−1, offsetting 9.7% of C emissions from transportation and energy sources. Accounting for multiple driving variables, we project a gradual decline in the forest C emission sink over the next 25 years (to 112u2009Tg yr−1) with regional differences. Sequestration in eastern regions declines gradually while sequestration in the Rocky Mountain region declines rapidly and could become a source of atmospheric C due to disturbances such as fire and insect epidemics. C sequestration in the Pacific Coast region stabilizes as forests harvested in previous decades regrow. Scenarios simulating climate-induced productivity enhancement and afforestation policies increase sequestration rates, but would not fully offset declines from aging and forest disturbances. Separating C transfers associated with land use changes from sequestration clarifies forests’ role in reducing net emissions and demonstrates that retention of forest land is crucial for protecting or enhancing sink strength.

Complex forest dynamics indicate potential for slowing carbon accumulation in the southeastern United States

Over the past century forest regrowth in Europe and North America expanded forest carbon (C) sinks and offset C emissions but future C accumulation is uncertain. Policy makers need insights into forest C dynamics as they anticipate emissions futures and goals. We used land use and forest inventory data to estimate how forest C dynamics have changed in the southeastern United States and attribute changes to land use, management, and disturbance causes. From 2007-2012, forests yielded a net sink of C because of net land use change (+6.48u2005Tg C yr−1) and net biomass accumulation (+75.4u2005Tg C yr−1). Forests disturbed by weather, insect/disease, and fire show dampened yet positive forest C changes (+1.56, +1.4, +5.48u2005Tg C yr−1, respectively). Forest cutting caused net decreases in C (−76.7u2005Tg C yr−1) but was offset by forest growth (+143.77u2005Tg C yr−1). Forest growth rates depend on age or stage of development and projected C stock changes indicate a gradual slowing of carbon accumulation with anticipated forest aging (a reduction of 9.5% over the next five years). Additionally, small shifts in land use transitions consistent with economic futures resulted in a 40.6% decrease in C accumulation.

The influence of multi-season imagery on models of canopy cover: A case study

Quantifying tree canopy cover in a spatially explicit fashion is important for broad-scale monitoring of ecosystems and for management of natural resources. Researchers have developed empirical models of tree canopy cover to produce geospatial products. For subpixel models, percent tree canopy cover estimates (derived from fi ne-scale imagery) serve as the response variable. The explanatory variables are developed from reflectance values and derivatives, elevation and derivatives, and other ancillary data. However, there is a lack of guidance in the literature regarding the use of leaf-on only imagery versus multi-season imagery for the explanatory variables. We compared models developed from leaf-on only Landsat imagery with models developed from multi-season imagery for a study area in Georgia. There was no statistical difference among models. We suggest that leaf-on imagery is adequate for the development of empirical models of percent tree canopy cover in the Piedmont of the Southeastern United States.

Watershed impacts of climate and land use changes depend on magnitude and land use context

Human population growth and urban development are affecting climate, land use, and the ecosystem services provided to society, including the supply of freshwater. We investigated the effects of land use and climate change on water resources in the Yadkin–Pee Dee River Basin of North Carolina, United States. Current and projected land uses were modeled at high resolution for three watersheds representing a forested to urban land use gradient by melding the National Land Cover Dataset with data from the U.S. Forest Service Forest Inventory and Analysis. Forecasts for 2051–2060 of regional land use and climate for scenarios of low (B2) and moderately high (A1B) rates of change, coupled with multiple global circulation models (MIROC, CSIRO, and Hadley), were used to inform a distributed ecohydrological model. Our results identified increases in water yields across the study watersheds, primarily due to forecasts of increased precipitation. Climate change was a more dominant factor for future water yield relative to land use change across all land uses (forested, urban, and mixed). When land use change was high (27% of forested land use was converted to urban development), it amplified the impacts of climate change on both the magnitude and timing of water yield. Our fine-scale (30-m) distributed combined modeling approach of land use and climate change identified changes in watershed hydrology at scales relevant for management, emphasizing the need for modeling efforts that integrate the effects of biophysical (climate) and social economic (land use) changes on the projection of future water resource scenarios.

Prediction of Canopy Heights over a Large Region Using Heterogeneous Lidar Datasets: Efficacy and Challenges

Generating accurate and unbiased wall-to-wall canopy height maps from airborne lidar data for large regions is useful to forest scientists and natural resource managers. However, mapping large areas often involves using lidar data from different projects, with varying acquisition parameters. In this work, we address the important question of whether one can accurately model canopy heights over large areas of the Southeastern US using a very heterogeneous dataset of small-footprint, discrete-return airborne lidar data (with 76 separate lidar projects). A unique aspect of this effort is the use of nationally uniform and extensive field data (~1800 forested plots) from the Forest Inventory and Analysis (FIA) program of the US Forest Service. Preliminary results are quite promising: Over all lidar projects, we observe a good correlation between the 85th percentile of lidar heights and field-measured height (r = 0.85). We construct a linear regression model to predict subplot-level dominant tree heights from distributional lidar metrics (R2 = 0.74, RMSE = 3.0 m, n = 1755). We also identify and quantify the importance of several factors (like heterogeneity of vegetation, point density, the predominance of hardwoods or softwoods, the average height of the forest stand, slope of the plot, and average scan angle of lidar acquisition) that influence the efficacy of predicting canopy heights from lidar data. For example, a subset of plots (coefficient of variation of vegetation heights <0.2) significantly reduces the RMSE of our model from 3.0–2.4 m (~20% reduction). We conclude that when all these elements are factored into consideration, combining data from disparate lidar projects does not preclude robust estimation of canopy heights.

Approximating prediction uncertainty for random forest regression models

Abstract Machine learning approaches such as random forest have increased for the spatial modeling and mapping of continuous variables. Random forest is a non-parametric ensemble approach, and unlike traditional regression approaches there is no direct quantification of prediction error. Understanding prediction uncertainty is important when using model-based continuous maps as inputs to other modeling applications such as fire modeling. Here we use a Monte Carlo approach to quantify prediction uncertainty for random forest regression models. We test the approach by simulating maps of dependent and independent variables with known characteristics and comparing actual errors with prediction errors. Our approach produced conservative prediction intervals across most of the range of predicted values. However, because the Monte Carlo approach was data driven, prediction intervals were either too wide or too narrow in sparse parts of the prediction distribution. Overall, our approach provides reasonable estimates of prediction uncertainty for random forest regression models.