Steen Magnussen

Sampling methods, remote sensing and GIS multiresource forest inventory

Forest Inventories - an Overview.- Forest Mensuration.- Sampling in Forest Surveys.- Remote Sensing.- Geographic and Forest Information Systems.- Multiresource Forest Inventory.

An accuracy assessment framework for large-area land cover classification products derived from medium-resolution satellite data

Land cover classification over large geographic areas using remotely sensed data is increasingly common as a result of the requirements of national inventory and monitoring programmes, scientific modelling and international environmental treaties. Although large‐area land cover products are more prevalent, standard operational protocols for their validation do not exist. This paper provides a framework for the accuracy assessment of large‐area land cover products and synthesizes some of the key decision points in the design and implementation of an accuracy assessment from the literature. The fundamental components of a validation plan are addressed and the framework is then applied to the land cover map of the forested area of Canada that is currently being produced by the Earth Observation for Sustainable Development programme. This example demonstrates the compromise between the theoretical aspects of accuracy assessment and the practical realities of implementation, over a specific jurisdiction. The framework presented in this paper provides an example for others embarking on the assessment of large‐area land cover products and can serve as the foundation for planning a statistically robust validation.

Contextual classification of Landsat TM images to forest inventory cover types

In the context of Landsat TM images forest stands are a cluster of homogeneous pixels. Contextual classification of forest cover types exploits relationships between neighbouring pixels in the pursuit of an increase in classification accuracy. Results with six contextual classifiers from two sites in Canada were compared to results with a maximum likelihood (ML) classifier. The comparisons were done at three levels of spectral class separation. Training and validation data were obtained from single-stage cluster sampling of 2 km×2 km primary sampling units (PSU) located on a 20 km×20 km grid. A strong relationship between contextual and ML classification accuracy was explored with logistic regression analysis. Effects of contextual classification were predicted for given levels of ML accuracy. Estimates of the spatial autocorrelation of reflectance values within a PSU were deemed consistent with a first-order autoregressive process. Iterative Conditional Modes (ICM) was the best contextual method; it improved the overall accuracy by four to six percentage points (statistically significant) when ML accuracy was between 50% and 80%. A relaxed ICM and a smoothing algorithm were second and third best. Contextual classification is most promising when an ML accuracy is around 70%. ICM results were sensitive to the level of spatial autocorrelation of ML classification errors and to the homogeneity of a PSU.

Use of models in large-area forest surveys: comparing model-assisted, model-based and hybrid estimation

This paper focuses on the use of models for increasing the precision of estimators in large-area forest surveys. It is motivated by the increasing availability of remotely sensed data, which facilitates the development of models predicting the variables of interest in forest surveys. We present, review and compare three different estimation frameworks where models play a core role: model-assisted, model-based, and hybrid estimation. The first two are well known, whereas the third has only recently been introduced in forest surveys. Hybrid inference mixes design-based and model-based inference, since it relies on a probability sample of auxiliary data and a model predicting the target variable from the auxiliary data..We review studies on large-area forest surveys based on model-assisted, model-based, and hybrid estimation, and discuss advantages and disadvantages of the approaches. We conclude that no general recommendations can be made about whether model-assisted, model-based, or hybrid estimation should be preferred. The choice depends on the objective of the survey and the possibilities to acquire appropriate field and remotely sensed data. We also conclude that modelling approaches can only be successfully applied for estimating target variables such as growing stock volume or biomass, which are adequately related to commonly available remotely sensed data, and thus purely field based surveys remain important for several important forest parameters.

A model-assisted k-nearest neighbour approach to remove extrapolation bias

Abstract In applications of the k-nearest neighbour technique (kNN) with real-valued attributes of interest (Y) the predictions are biased for units with ancillary values of X with poor or no representation in a sample of n units. In this article a model-assisted calibration is proposed that reduces unit-level extrapolation bias. The bias is estimated as the difference in model-based predictions of Y given the X-values of the true k nearest units and the k selected reference units. Calibrated kNN predictions are then obtained by adding this difference to the original kNN prediction. The relationship is modelled between Y and X with decorrelated X-variables, variables scaled to the interval [0,1] and Bernstein basis functions to capture changes in Y as a function of changes in X. Three examples with actual forest inventory data from Italy, the USA and Finland demonstrated that calibrated kNN predictions were, on average, closer to their true values than non-calibrated predictions. Calibrated predictions had a range much closer to the actual range of Y than non-calibrated predictions.

Curve fitting of time-series Landsat imagery for characterizing a mountain pine beetle infestation

In this technical note we present a new technique using mixed linear models for characterizing a mountain pine beetle (Dendroctonus ponderosae Hopkins) infestation from multiyear satellite imagery. The main benefit of our approach is an ability to determine the statistical significance of each annual spectral change. Knowledge of the annual spectral change characteristics can then be used to statistically determine if a disturbance event has occurred, the timing of a given disturbance event, as well as to provide information for clustering fitted multitemporal reflectance curves (i.e. spectral trajectories) with a common shape. The spatial clustering of spectral trajectories provides insights into the nature of the disturbance and recovery imposed by infestation over a 14-year period.

Two new density estimators for distance sampling

Two new density estimators for k-tree distance sampling are proposed and their performance is assessed in simulated distance sampling from 22 stem maps representing a wide range of natural to semi-natural forest tree stands with random to irregular (clustered) spatial distribution of trees. The new estimators are model-based. The first (Orbit) computes density as the inverse of the average of the areas associated with each of the k-trees nearest to a sample location. The area of the k-th tree is obtained as a prediction from a linear regression model while the area of the first is obtained via a Poisson probability integral. The second (GamPoi) is based on the expected distribution of distance to the k nearest tree in a forest where the local distribution of trees is random but the stem density varies from sample location to sample location as a gamma distribution. In a comprehensive assessment with 17 promising reference estimators, a subset composed of Morisita’s, Persson’s, Byth’s, Kleinn’s, Orbit, and GamPoi was significantly better, in terms of relative root mean square error (RRMSE), than average. GamPoi emerged as the better estimator for sample sizes larger than or equal to 30. For smaller sample sizes, both Kleinn’s and Morisita’s appear attractive.

Selecting Representative High Resolution Sample Images for Land Cover Studies. Part 2: Application to Estimating Land Cover Composition

Abstract We tested the effectiveness of the Purposive Selection Algorithm (PSA, described in the companion first article) to accurately estimate land cover composition over a large area. The knowledge of land cover distribution over large areas is increasingly more important for numerous scientific and policy purposes. Unless a complete detailed map is necessary, a sampling approach is the best strategy for determining the relative proportions of individual cover types because of its cost-effectiveness and speed of application. With coarse resolution land cover maps at continental or global scales increasingly becoming available, the possibility arises of using such maps synergistically with a sample of high resolution satellite coverage. The goal of such synergy would be to obtain accurate estimates of land cover composition over large areas as well as the knowledge of local spatial distribution. We evaluated PSA performance for sample selection over a 136,432 km 2 area (domain) in the BOREAS Region of Saskatchewan and Manitoba, Canada. Two maps were prepared for the domain, one based on NOAA Advanced Very High Resolution Radiometer (AVHRR, 1 km pixels) and one on LANDSAT Thematic Mapper (TM, 30 m). After dividing the area into 134 tiles, a PSA sample was selected using the AVHRR tiles. A random sample was also selected for comparison. The domain AVHRR cover type fractions were then corrected using TM maps for the selected tiles, following the method of Walsh and Burk (1993) . The land cover composition obtained through the combined “domain AVHRR/sample TM” data was then compared with the domain TM coverage. We found that PSA provided a representative sample to correct the AVHRR map, particularly for small sample sizes. Compared to the random selection, PSA yielded more accurate results at all tested sampling fractions (up to 30% of all tiles). With a PSA sample of 7% (18%), the average absolute difference per class between the correct and the estimated fraction was 0.058% (0.043%). For the same sample fractions, the average relative error per class was 16.1% (9.8%) for PSA and 24.5% (18.7%) for random selection. The difference between PSA and random selections was significant at the 0.001 probability level. It is concluded that the PSA strategy is an effective way to combine coarse and fine resolution satellite data to obtain expedient and cost-effective land cover information over large areas. An important benefit of the synergistic combination of the two maps is knowledge of land cover distribution at the landscape level. This is because the coarse resolution map provides the overall distribution patterns across the domain, while the fine resolution map supplies the average composition of the coarse resolution pixels in each cover type. Thus, each coarse pixel can be statistically divided into the component high resolution classes. Crown

Prediction of daily lightning- and human-caused fires in British Columbia

Daily records of the location and timing of human- and lightning-caused fires in British Columbia from 1981 to 2000 were used to estimate the probability of fire occurrence within 950 20 × 20-km spatial units (~950 000 km2) using a binary logistic regression modelling framework. Explanatory variables included lightning strikes, forest cover, surface weather observations, atmospheric stability indices and fuel moisture codes of the Canadian Fire Weather Index System. Because the influence of the explanatory variables in the models varied from year to year, model coefficients were estimated for each year. The arithmetic mean of the model coefficients was used for making daily predictions in a future year. A confidence interval around the mean or a quantile was derived from the ensemble of 20 model predictions. A leave-1-year-out cross-validation procedure was used to assess model performance for random years. The daily number of lightning-caused fires was reasonably well predicted at the provincial level (R = 0.83) and slightly less well predicted for a smaller (75 000 km2) administrative region. The daily number of human-caused fires was less well predicted at both the provincial (R = 0.55) and the regional level. The ability to estimate confidence intervals from the ensemble of model predictions is an advantage of the year-specific approach.

Post-Fire Canopy Height Recovery in Canada’s Boreal Forests Using Airborne Laser Scanner (ALS)

Canopy height data collected with an airborne laser scanner (ALS) flown across unmanaged parts of Canadas boreal forest in the summer of 2010 were used—as stand-alone data—to derive a least-squares polynomial (LSPOL) between presumed post-fire recovered canopy heights and duration (in years) since fire (YSF). Flight lines of the >25,000-km ALS survey intersected 163 historic fires with a known day of detection and fire perimeter. A sequential statistical testing procedure was developed to separate post-fire recovered canopy heights from pre-fire canopy heights. Of the 153 fires with >5 YSF, 121 cases (89%) could be resolved to a complete or partial post-fire canopy replacement. The estimated LSPOL can be used to estimate post-fire aboveground biomass and carbon sequestration in areas where alternative information is dated or absent. These LIDAR derived findings are especially useful as existing growth information is largely developed for higher productivity ecosystems and not applicable to these ecosystems subject to large wildfires.