Assessing forest condition from airborne remotely sensed data

Authors: Hill, R., Sumnall, M.J. and Hinsley, S.A.

Conference: ForestSAT 2014: Operational Tools in Forestry Using Remote Sensing Techniques

Dates: 4-7 November 2014

Abstract:

The assessment of condition or quality is of key importance to the management of any woodland area. However, the concept of condition is difficult to define and objectify. For example, there are many differences in the definition of woodland condition between research, conservation and commercial interests; with factors such as species richness or biodiversity, ecosystem health or resilience, naturalness and turnover, structure and productivity, all potentially considered as condition determinants. Defining woodland condition in terms of an organism or community (perhaps of a keystone or indicator species) is one approach, but this can result in an ecologically restricted or biased assessment. Often, to ascertain condition in a woodland ecosystem necessitates a specific aim (i.e. desired condition) to be identified, thereby introducing the idea of targets which themselves should be measurable and accurately quantifiable. For an individual forest patch, such as a compartment or stand, condition indicators are typically grouped according to structure, composition, regeneration, turnover (i.e. deadwood), and health (including damage or disease). By focussing within a patch, this effectively eliminates factors relating to landscape two-dimensional patch metrics.

This paper presents a sensor-fusion approach to surveying a forested field site, with the development of habitat condition indicators derived from remotely sensed data. Estimation of forest compositional and structural characteristics was made from airborne hyperspectral data and small footprint LiDAR capturing both discrete-return (DR) and full-waveform (FW) data.

This combination of remote sensing data offered detailed and complimentary information about the structure and composition of woodlands.

A 22 km2 study area was established in the New Forest, Hampshire, UK (Lat. 50.84° N, Long. 1.50° W) which contained a variety of forest structural types, including managed plantation, and semi-ancient coniferous and deciduous woodland. Field data collection took place in 41 field plots located across a stratification of forest structural properties. A total of 37 forest variables were recorded in each of the 30x30 m sized field plots. These variables covered aspects of forest structure (14 variables), composition (7 variables), deadwood (4 variables), understorey and ground vegetation (12 variables).

The hyperspectral and LiDAR data were acquired simultaneously under both leaf-off and leaf-on conditions in 2010 (April and July respectively). Various metrics were extracted from each dataset corresponding with the field plot areas and these were used in statistical analyses in order to estimate field plot-level forest metrics, taking both an Ordinary Least Squares (OLS) linear regression and Akaike's Information Criterion (AIC) approach. LiDAR metrics such as descriptive statistics, skewness, kurtosis, and height percentiles were extracted from the point clouds of leaf-on and leaf-off data, classified as vegetation or ground returns, for DR and FW separately. In addition, a further 8 metrics relating to crown geometry and separation were extracted from an individual tree crown (ITC) database created from the leaf-on DR and FW lidar data separately. Hyperspectral metrics were extracted from 13 spectral indices (generated from leaf-on and leaf-off data separately), plus an additional 6 variables relating to species diversity derived from an ITC-based tree species map.

The most accurate models were selected from each of the remote sensing datasets, either separately or in combination. From the selection of best models, 32 of the 37 field plot-level (30x30m) forest variables could be estimated with an NRMSE below 0.4. These forest variables could therefore be mapped relatively successfully across the study site and used in combination in a series of forest condition assessment methods.

Six conventional forest condition assessment indices were assessed, of which three were successful, corresponding with independent field validation data and forest compartment boundaries. These assessment methods (complexity index, complex stand diversity index, and a score-based index) were driven primarily by tree species and tree size variables. In addition for the score-based method, the presence of deadwood was also important (found to be positively linked to increases in tree size), and understorey composition (found to be independent from tree size variables). The best technique for assessing woodland condition was the score-based method which combined seventeen inputs relating to tree species composition, tree size and variability, deadwood, and understory components.

This approach demonstrates that conventional methods of assessing forest condition can be applied with remote sensing derived inputs for woodland assessment purposes over landscape-scale areas. While the requirement for fieldwork remains, the approach demonstrated here provides a far more detailed and/or comprehensive indication of condition than is possible through field work assessment alone in terms of spatial extent, and on a scale appropriate to observe spatial patterns of features within the stand-level.

Source: Manual