P3-3: Scale Mismatch: A Major Challenge in Integrated Forest Monitoring, Combining Remote Sensing and Sample Based Field Observations

PhD student: Wanda Graf
Supervisor: Prof. Dr. Christoph Kleinn
Group: Forest Inventory and Remote Sensing

Project Description:
Forest monitoring has the overall goal to generate a reliable base of data and of meaningful information for informed decision making in forest management and forest policy, but also for forest related research. Scaling issues are "omnipresent" in forestry and forest monitoring, both in terms of temporal and spatial scales. Here, we focus on spatial scales within the domain that is described in the following.
Forest monitoring builds on two major data sources (and as in all empirical observational studies these require clear definitions of variables and observation/measurement approaches): sample based field observations and remotely sensed observations. Both carry inherent scaling issues which pose a scientific challenge both for estimation and for integration, but also for interpretation of results: (1) field sampling in the forest is done on sample plots of variable size, shape and type. The underlying population of sample plots is described with the so-called "infinite population approach" that looks at the set of all points in the study area as the population to be sampled; the sample trees observed around that point are considered "support". Depending on size, shape and type of observation plots, however, this population varies - even though the underlying forest area is the same. This scale issue in observation design, obviously, affects precision of estimation and also the matching to remotely sensed image information. (2) In remote sensing, different imagery has different resolution, including spatial resolution, and carries issues in matching them to the field observations; in addition, spectral resolution varies: the type of observations that can be made and the possibility to distinguish different features depends critically on the specific combination of spectral ranges.
The underlying problem of the scale of measurement/observation is a general one and relevant in many fields in particular in ecology: estimation of vegetation cover, for example, depends on the scale of observation (Magdon and Kleinn 2013) as does the estimation of forest edge length and other measures of landscape fragmentation (Kleinn et al. 2011).
For this and alike problems, the term "modifiable areal unit problem" (MAUP) has been coined which states that the inference may change with the shape and scale of observational units. The MAUP is one special aspect of the more general "change of support problem" (COSP), i.e., combining spatially (incompatible) data across different scales. Potential solutions to the COSP have received increasing attention, but their applicability in the context of forest monitoring has raredly been tested.
Information on forest resources is not only needed at a population level, i.e., entire forests, but also for domains such as forest types, individual stands, etc. When only few or no field plot observations are available at the domain level, direct estimation of domain properties becomes unfeasible. A common approach is to "borrow strength" from outside the area of interest using mixed models. In forest inventory applications remote sensing data is used as auxiliary information in these models. The effect of the COSP on parameter estimates may differ whether estimates are demanded for populations or (sub-) domains.
In this project, two basic methodological issue of integrated forest monitoring shall be investigated: (1) We will test whether the proposed solutions to the COSP are applicable to data from forest resource assessments. Here, we are particularly interested in combining high spatial, and spectral resolution imagery and Airborne Laser Scanning (ALS) data with field plots of different size and shape. (2) We will investigate the effect of the COSP on small-area estimation, i.e., estimation of inventory target parameters like biomass, timber volume or tree species distribution for individual forest stands. An additional challenge comes in because field observations (plots) and remotely sensed observations can hardly be perfectly geographically matched; there is always a so-called co-registration error, that is: a mismatch of unknown size which - depending on spatial autocorrelation and the design characteristics of the plots - will reduce precision of estimation.
Several data sets of geo-referenced field observations and remote sensing imagery of different resolutions are available from other projects in the forests around Göttingen; additional imagery and additional field work (research students) will be necessary to complement the available data sets.