This dissertation addresses critical aspects of these questions by: quantifying and mitigating the impact of co-registration errors; comparing various sample designs and estimation tech
Estimates of forest characteristics derived from remotely sensed imagery and field samples: applicable scales, appropriate study design, and relevance to forest management
Accurate forest information is essential for effective management, with stand metrics such as species composition, basal area (BAH, m2/ha), and trees per hectare (TPH) guiding decisions Collecting these metrics at fine spatial detail across large landscapes via ground surveys alone is cost-prohibitive, prompting the development of a remotely sensed methodology to quantify BAH and TPH at fine to medium spatial resolutions over broad extents The work addresses theoretical concerns—spatial scale, co-registration errors, optimal field sampling unit configurations, sampling intensity and allocation, and the derivation of BAH and TPH estimates—and applies them to the northwest Florida region known as the Florida Panhandle By linking field measurements to high-resolution remotely sensed data, patterns of BAH and TPH are quantified across broad extents, enabling more efficient, decision-relevant natural resource management The primary objectives are to close knowledge gaps across forestry, remote sensing, data science, and decision science so findings can inform management decisions at fine spatial detail across large areas.
Keywords: basal area, trees density, co-registration, sample design, longleaf, forest characteristics
Forest management is a complex, integrated process that aims to balance economic, ecological, and social objectives for forested lands Since the National Forest Management Act of 1976, the federal definition of forest management has broadened beyond timber production to include economic and social goals as core components of management decisions, the consideration of broad, socially defined multiple-use management problems, and the requirement to provide quantitative justification for forest plans and choices This expanded scope changes not only what we manage for, but how we justify our decisions, placing greater emphasis on planning that spans larger spatial extents and broader contextual considerations.
Across diverse forests with varying ownership, complexity, and size, forest plans guide management activities and silvicultural prescriptions toward private, public, and broader social objectives Effective planning and implementation require knowledge of both biotic and abiotic forest conditions and an understanding of how these conditions interact within the defined goals for a given forest To assess the current structure and composition, practitioners rely on established mensuration techniques that aggregate data from field plots across a geographic area to estimate mean values and variances of forest characteristics While these methods are well described, they can be costly and challenging to apply at fine spatial resolutions over large extents As the human population grows and more people rely on forested landscapes, social questions about the impacts of management on forest ecosystems, connectivity, and long-term sustainability come to the fore.
Water quality, esthetics, carbon, air quality, climate, and timber products markets are increasingly influenced by finer spatial detail The high cost of quantifying the basic information needed to describe these forest characteristics means that relevant data are often scarce at the spatial resolution required for management decisions As a result, decision-makers frequently contend with limited information at the implementation scale, which can challenge effective planning and optimization of forest resources.
Well-known inventory programs, such as the Forest Inventory Analysis Program of the U.S Forest Service, provide a wealth of data on the nation’s forests; however, the inferences drawn from these data are most reliable at regional spatial resolutions and offer limited utility at national or state scales This reflects a mismatch between the intent and scope of data collection and society’s needs for spatially explicit forest-management insights, meaning the data are not useless but do not directly address broad spatial questions To support projects like timber cruises or sales, forest managers must implement more intensive sampling schemes; yet these efforts tend to be inconsistent, vary in intensity and scope, and cover only small geographic areas (often under 1,000 hectares), making the collected data incongruent with other inventory efforts and impractical to deploy at broad extents.
To illustrate the financial constraints of intensive field plot inventories, it is useful to compare per-plot costs across approaches The FIA protocol for collecting basic forest information carries an estimated cost of $600 to $1,240 per plot, highlighting the higher investment required for comprehensive data By contrast, timber cruises conducted primarily to estimate timber volume can incur substantially lower per-plot costs, sometimes only a few hundred dollars, underscoring the broad cost spectrum among field inventory methods.
Plot costs ranging from $50 to $1,240 imply that a 10% cruise over 100,000 hectares would require about 100,000 plots, each with a radius of 11.3 m, totaling between $5 million (at $60 per plot) and $124 million (at $1,240 per plot) Even if a 10% sample were deemed sufficient to capture forest complexity for planning, such an expenditure makes broad-scale, high-resolution assessment prohibitive, so practitioners often resort to coarse representations that describe forest characteristics as totals or averages within defined areas, which in turn forces planners to develop generalized forest plans with significant uncertainty about fine-scale forest conditions Forest traits such as species composition, spatial arrangement, basal area (BAH, m2 ha-1), and tree densities (TPH, trees ha-1), while not inherently costly to measure at the plot level, become expensive because of the large number of plots required to quantify stand characteristics across a stand or strata, since classical inventory methods split the forest into stands of similar composition, stocking, size, and age class and then summarize plots within each stand to estimate the mean and variance of BAH and TPH for informing the planning process.
Although this procedure can be applied in almost any situation and requires only plot data, it has been widely adopted in forestry and typically yields estimates for the stand as a whole, usually requiring a large sample size and does not directly allow for additional sources of information When additional forest information is known, the classical approach has been expanded to include that information by grouping stands into like strata Stratification aims to reduce sampling variation within like groups (the strata), thereby reducing sampling intensity and cost to achieve a predefined level of accuracy Within each stratum, plot data are summarized, and the mean and variance for a given variable are attributed to stands and then pooled or combined in a weighted fashion to estimate an overall mean and variance for the forest as a whole.
When supplementary information—such as remotely sensed data—about the population (e.g., BAH) is available and correlated with the population variable of interest, regression can be used to further improve the precision and efficiency of a given sample In this estimation framework, supplemental information can be categorical or continuous, is tested for relevance in terms of minimizing variation, and is used to estimate the strength of its relationship with the target variable.
4 between the response variable (e.g., BAH) and predictor variables (e.g spectral values from imagery) While regression has been used by biometricians to develop many allometric equations [11], this technique has only recently been used on a limited basis to estimate key stand metrics such as species composition, BAH, and TPH for a forest Historically, this may have been due to the availability, scale, and quality of supplemental information with regard to plots and stands within a forest Today, however, there is a wealth of digital and remotely sensed data (e.g., [12-15]) that can be used to increase the precision of estimates of key stand metrics used to inform forest management, while simultaneously reducing sampling cost
For many years, remotely sensed data have been used to explore our surroundings [16] and stratify the terrestrial environment in useful ways [17, 18] With recent advancements in technology, mathematics, statistics, machine learning, and computer science, remotely sensed relationships between reflected portions of the electromagnetic spectrum and the earth’s terrestrial surface have been documented and exploited to build a wide range of data products depicting terrestrial characteristics such as topography [19], land use and cover [20], vegetative indices [21], vegetation communities [22, 23], fire severity [24], land cover change [25], and temperature [26] Spatially defining these terrestrial characteristics has elevated the importance of fields such as landscape ecology [27] in understanding the impacts of patterns within a forest as they relate to the landscape- scale functions and services they provide Within the context of forest management, these concepts underlie the necessity of accurately quantifying not only general amounts or resources and the condition of the forest as a whole, but spatially depicting spatial variations in forest characteristics such as species BAH and TPH with a high degree of fidelity
Despite the clear gains in sampling efficiency and the accuracy of forest-characteristics estimates achieved with regression techniques, along with the wealth of complementary remotely sensed data now available, regression has not been fundamentally adopted to quantify metrics such as BAH and TPH at plot-, stand-, and forest-scale levels This limited uptake stems from practical barriers, notably a lack of familiarity with remotely sensed data, among other challenges.
Historically, remotely sensed data have shown coarse spatial resolution, limiting fine-scale forest analysis; this, along with technical challenges in modeling and processing the data, hampers effective interpretation The acquisition of remote sensing data also incurs additional costs, which can be a barrier for many projects Compounding these issues is the lack of robust statistical techniques and the weak or nonexistent statistical relationships between coarse remote-sensing metrics and traditional forestry indicators, complicating the integration of remote sensing into forestry assessments.
Field measurements of BAH and TPH have been successfully related to fine-grained remotely sensed data, enabling predictive models to create surfaces that predict BAH and TPH continuously across forests at plot-level resolution, with cell estimates aggregating to stands and forests Building on these spatially explicit outputs, collaborations have yielded techniques to optimize sustained yield across a 202,000‑hectare forested landscape and to estimate delivered costs and feedstock supply for more than 8 million hectares Together, these results illustrate that forest characteristics derived from linking field plots to remotely sensed data provide the baseline characterization of both stands and the broader forest needed to perform various fine-resolution, spatially explicit analyses.
2 Summary of the Chapter Contributions
While education and outreach continue to address many practical issues of remotely sensed data, unresolved questions persist about scale, sample design, modeling approaches, and the usefulness of derived outputs for forest planning and management, even as new fine-grained sensors such as Sentinel-2 and advances in processing techniques and software emerge This dissertation tackles these issues from theoretical and applied perspectives, drawing on the tenets of data and decision science In chapters 2 and 3, the work quantifies co-registration errors and demonstrates how their impacts can be minimized through spatial aggregation, and outlines the benefits of sample designs that spread and balance observations across predictor-variable space for various estimation tasks.
Mitigating the Impact of Field and Image Registration Errors through Spatial Aggregation
Remotely sensed data are widely used as predictor variables in spatially explicit models that describe landscape characteristics across broad extents at relatively fine resolution Building these models requires spatially registering predictor data to a known coordinate system and linking responses to predictor values, a process that inherently introduces measurement error into both the response and the predictors, with the latter causing attenuation bias Through simulations, we show that the spatial correlation between the response and predictor fields and their co-registration errors can substantially affect bias and the accuracy of linear models The study also evaluates spatial aggregation as a strategy to mitigate co-registration effects, examines subsampling within the extent of sampling units, and provides a technique to determine the observational unit size needed to minimize co-registration impact while quantifying the potential error in predictive models.
Remotely sensed data play an ever-increasing role in characterizing and quantifying landscapes These types of data have been used to study our surroundings [1], stratify the terrestrial environment
[2, 3], and build a wide range of data products depicting terrestrial characteristics, such as topography
Key data layers include land use and land cover, vegetative indices, vegetation communities, fire severity, land-cover change, and temperature Because remotely sensed data have proven effective and relatively inexpensive for depicting landscape patterns and their changes over time, remote sensing has become foundational in landscape ecology.
[12] and concepts like spatial connectivity and the relationships between patterns and processes are now at the forefront of many land management and planning endeavors [13–16]
Concepts like spatial contiguity, patch size, and patch juxtaposition, and their links to forest management, land use planning, and sustainable forestry have driven the effort to define landscape patterns with fine spatial detail across broad extents With access to fine-grained remotely sensed data (