Murray Richardson

On the Importance of Training Data Sample Selection in Random Forest Image Classification: A Case Study in Peatland Ecosystem Mapping

Random Forest (RF) is a widely used algorithm for classification of remotely sensed data. Through a case study in peatland classification using LiDAR derivatives, we present an analysis of the effects of input data characteristics on RF classifications (including RF out-of-bag error, independent classification accuracy and class proportion error). Training data selection and specific input variables (i.e., image channels) have a large impact on the overall accuracy of the image classification. High-dimension datasets should be reduced so that only uncorrelated important variables are used in classifications. Despite the fact that RF is an ensemble approach, independent error assessments should be used to evaluate RF results, and iterative classifications are recommended to assess the stability of predicted classes. Results are also shown to be highly sensitive to the size of the training data set. In addition to being as large as possible, the training data sets used in RF classification should also be (a) randomly distributed or created in a manner that allows for the class proportions of the training data to be representative of actual class proportions in the landscape; and (b) should have minimal spatial autocorrelation to improve classification results and to mitigate inflated estimates of RF out-of-bag classification accuracy.

Wetland mapping with LiDAR derivatives, SAR polarimetric decompositions, and LiDAR–SAR fusion using a random forest classifier

In this paper, we assess the use of Random Forest (RF) for mapping land cover classes within Mer Bleue bog, a large northern peatland in southeastern Ontario near Ottawa, Canada, using Synthetic Aperture Radar (SAR) and airborne Light Detection and Ranging (LiDAR). Not only has RF been shown to improve classification accuracies over more traditional classifiers, but it also provides useful information on the statistical importance of individual input image bands for land cover classification. Our specific objectives in this study were to: (i) assess the robustness of a RF approach to northern peatland classification; (ii) examine variable importance resulting from the RF classifications to identify which imagery types, derivatives, and analysis scales are most useful for mapping different classes of northern peatlands; (iii) assess if fusion of different LiDAR and SAR variables can improve classification accuracies at Mer Bleue; and (iv) assess physical interpretability of the multisensor image types and derivatives with respect to biophysical attributes associated with peatland classes. Our results show that the fusion of SAR with LiDAR imagery and derivatives at this study site did not provide additional classification accuracy over the use of LiDAR derivatives alone. Nevertheless, the RF-based approach presented here has strong potential to improve mapping and imagery classification of wetlands and may also help researchers and practitioners improve information extraction and land cover classification in other application areas benefitting from large volumes of multi-sensor imagery.

The role of terrestrial vegetation in atmospheric Hg deposition: Pools and fluxes of spike and ambient Hg from the METAALICUS experiment

[1] As part of the Mercury Experiment to Assess Atmospheric Loading in Canada and the U.S. (METAALICUS), different stable Hg(II) isotope spikes were applied to the upland and wetland areas of a boreal catchment between 2001 and 2006 to examine retention of newly deposited Hg(II). In the present study, a Geographical Information Systems (GIS)-based approach was used to quantify canopy and ground vegetation pools of experimentally applied upland and wetland spike Hg within the METAALICUS watershed over the terrestrial loading phase of the experiment. A chemical kinetic model was also used to describe the changes in spike Hg concentrations of canopy and ground vegetation over time. An examination of the fate of spike Hg initially present on canopy vegetation using a mass balance approach indicated that the largest percentage flux from the canopy over one year post-spray was emission to the atmosphere (upland: 45%; wetland: 71%), followed by litterfall (upland: 14%; wetland: 10%) and throughfall fluxes (upland: 12%; wetland: 9%) and longer term retention of spike in the forest canopy (11% for both upland and wetland). Average half-lives (t1/2) of spike on deciduous (110 � 30 days) and coniferous (180 � 40 days) canopy and ground vegetation (890 � 620 days) indicated that retention of new atmospheric Hg(II) on terrestrial (especially ground) vegetation delays downward transport of new atmospheric Hg(II) into the soil profile and runoff into lakes.

Assessing the Potential to Operationalize Shoreline Sensitivity Mapping: Classifying Multiple Wide Fine Quadrature Polarized RADARSAT-2 and Landsat 5 Scenes with a Single Random Forest Model

The Random Forest algorithm was used to classify 86 Wide Fine Quadrature Polarized RADARSAT-2 scenes, five Landsat 5 scenes, and a Digital Elevation Model covering an area approximately 81,000 km2 in size, and representing the entirety of Dease Strait, Coronation Gulf and Bathurst Inlet, Nunavut. The focus of this research was to assess the potential to operationalize shoreline sensitivity mapping to inform oil spill response and contingency planning. The impact of varying the training sample size and reducing model data load were evaluated. Results showed that acceptable accuracies could be achieved with relatively few training samples, but that higher accuracies and greater probabilities of correct class assignment were observed with larger sample sizes. Additionally, the number of inputs to the model could be greatly reduced without impacting overall performance. Optimized models reached independent accuracies of 91% for seven land cover types, and classification probabilities between 0.77 and 0.98 (values for latter represent per-class averages generated from independent validation sites). Mixed results were observed when assessing the potential for remote predictive mapping by simulating transferability of the model to scenes without training data.

Fusion of Multispectral Imagery and Spectrometer Data in UAV Remote Sensing

Abstract: High spatial resolution hyperspectral data often used in precision farming applications are not available from current satellite sensors, and difficult or expensive to acquire from standard aircraft. Alternatively, in precision farming, unmanned aerial vehicles (UAVs) are emerging as lower cost and more flexible means to acquire very high resolution imagery. Miniaturized hyperspectral sensors have been developed for UAVs, but the sensors, associated hardware, and data processing software are still cost prohibitive for use by individual farmers or small remote sensing firms. This study simulated hyperspectral image data by fusing multispectral camera imagery and spectrometer data. We mounted a multispectral camera and spectrometer, both being low cost and low weight, on a standard UAV and developed procedures for their precise data alignment, followed by fusion of the spectrometer data with the image data to produce estimated spectra for all the multispectral camera image pixels. To align the data collected from the two sensors in both the time and space domains, a post-acquisition correlation-based global optimization method was used. Data fusion, to estimate hyperspectral reflectance, was implemented using several methods for comparison. Flight data from two crop sites, one being tomatoes, and the other corn and soybeans, were used to evaluate the alignment procedure and the data fusion results. The data alignment procedure resulted in a peak R2 between the spectrometer and camera data of 0.95 and 0.72, respectively, for the two test sites. The corresponding multispectral camera data for these space and time offsets were taken as the best match to a given spectrometer reading, and used in modelling to estimate hyperspectral imagery from the multispectral camera pixel data. Of the fusion approaches evaluated, principal component analysis (PCA) based models and Bayesian imputation reached a similar accuracy, and outperformed simple spline interpolation. Mean absolute error (MAE) between predicted and observed spectra was 17% relative to the mean of the observed spectra, and root mean squared error (RMSE) was 0.028. This approach to deriving estimated hyperspectral image data can be applied in a simple fashion at very low cost for crop assessment and monitoring within individual fields.

Moving to the RADARSAT Constellation Mission: Comparing Synthesized Compact Polarimetry and Dual Polarimetry Data with Fully Polarimetric RADARSAT-2 Data for Image Classification of Peatlands

For this research, the Random Forest (RF) classifier was used to evaluate the potential of simulated RADARSAT Constellation Mission (RCM) data for mapping landcover within peatlands. Alfred Bog, a large peatland complex in Southern Ontario, was used as a test case. The goal of this research was to prepare for the launch of the upcoming RCM by evaluating three simulated RCM polarizations for mapping landcover within peatlands. We examined (1) if a lower RCM noise equivalent sigma zero (NESZ) affects classification accuracy, (2) which variables are most important for classification, and (3) whether classification accuracy is affected by the use of simulated RCM data in place of the fully polarimetric RADARSAT-2. Results showed that the two RCM NESZs (−25 dB and −19 dB) and three polarizations (compact polarimetry, HH+HV, and VV+VH) that were evaluated were all able to achieve acceptable classification accuracies when combined with optical data and a digital elevation model (DEM). Optical variables were consistently ranked to be the most important for mapping landcover within peatlands, but the inclusion of SAR variables did increase overall accuracy, indicating that a multi-sensor approach is preferred. There was no significant difference between the RF classifications which included RADARSAT-2 and simulated RCM data. Both medium- and high-resolution compact polarimetry and dual polarimetric RCM data appear to be suitable for mapping landcover within peatlands when combined with optical data and a DEM.

Random forests as cumulative effects models: A case study of lakes and rivers in Muskoka, Canada

Cumulative effects assessment (CEA) - a type of environmental appraisal - lacks effective methods for modeling cumulative effects, evaluating indicators of ecosystem condition, and exploring the likely outcomes of development scenarios. Random forests are an extension of classification and regression trees, which model response variables by recursive partitioning. Random forests were used to model a series of candidate ecological indicators that described lakes and rivers from a case study watershed (The Muskoka River Watershed, Canada). Suitability of the candidate indicators for use in cumulative effects assessment and watershed monitoring was assessed according to how well they could be predicted from natural habitat features and how sensitive they were to human land-use. The best models explained 75% of the variation in a multivariate descriptor of lake benthic-macroinvertebrate community structure, and 76% of the variation in the conductivity of river water. Similar results were obtained by cross-validation. Several candidate indicators detected a simulated doubling of urban land-use in their catchments, and a few were able to detect a simulated doubling of agricultural land-use. The paper demonstrates that random forests can be used to describe the combined and singular effects of multiple stressors and natural environmental factors, and furthermore, that random forests can be used to evaluate the performance of monitoring indicators. The numerical methods presented are applicable to any ecosystem and indicator type, and therefore represent a step forward for CEA.

A Systematic Approach for Variable Selection With Random Forests: Achieving Stable Variable Importance Values

Random Forests variable importance measures are often used to rank variables by their relevance to a classification problem and subsequently reduce the number of model inputs in high-dimensional data sets, thus increasing computational efficiency. However, as a result of the way that training data and predictor variables are randomly selected for use in constructing each tree and splitting each node, it is also well known that if too few trees are generated, variable importance rankings tend to differ between model runs. In this letter, we characterize the effect of the number of trees (ntree) and class separability on the stability of variable importance rankings and develop a systematic approach to define the number of model runs and/or trees required to achieve stability in variable importance measures. Results demonstrate that both a large ntree for a single model run, or averaged values across multiple model runs with fewer trees, are sufficient for achieving stable mean importance values. While the latter is far more computationally efficient, both the methods tend to lead to the same ranking of variables. Moreover, the optimal number of model runs differs depending on the separability of classes. Recommendations are made to users regarding how to determine the number of model runs and/or trees that are required to achieve stable variable importance rankings.

Ratio of Methylmercury to Dissolved Organic Carbon in Water Explains Methylmercury Bioaccumulation Across a Latitudinal Gradient from North-Temperate to Arctic Lakes

We investigated monomethylmercury (MMHg) bioaccumulation in lakes across a 30° latitudinal gradient in eastern Canada to test the hypothesis that climate-related environmental conditions affect the sensitivity of Arctic lakes to atmospheric mercury contamination. Aquatic invertebrates (chironomid larvae, zooplankton) provided indicators of MMHg bioaccumulation near the base of benthic and planktonic food chains. In step with published data showing latitudinal declines in atmospheric mercury deposition in Canada, we observed lower total mercury concentrations in water and sediment of higher latitude lakes. Despite latitudinal declines of inorganic mercury exposure, MMHg bioaccumulation in aquatic invertebrates did not concomitantly decline. Arctic lakes with greater MMHg in aquatic invertebrates either had (1) higher water MMHg concentrations (reflecting ecosystem MMHg production) or (2) low water concentrations of MMHg, dissolved organic carbon (DOC), chlorophyll, and total nitrogen (reflecting lake sensitivity). The MMHg:DOC ratio of surface water was a strong predictor of lake sensitivity to mercury contamination. Bioaccumulation factors for biofilms and seston in Arctic lakes showed more efficient uptake of MMHg in low DOC systems. Environmental conditions associated with low biological production in Arctic lakes and their watersheds increased the sensitivity of lakes to MMHg.

Delineation of peatland lagg boundaries from airborne LiDAR

In Canada, peatlands are the most common type of wetland, but boundary delineation in peatland complexes has received little attention in the scientific literature. Typically, peatland boundaries are mapped as crisp, absolute features, and the transitional lagg zone—the ecotone found between a raised bog and the surrounding mineral land—is often overlooked. In this study, we aim (1) to advance existing approaches for detecting and locating laggs and lagg boundaries using airborne LiDAR surveys and (2) to describe the spatial distribution of laggs around raised bog peatlands. Two contrasting spatial analytical approaches for lagg detection were tested using five LiDAR-derived topographic and vegetation indices: topography, vegetation height, topographic wetness index, the standard deviation of the vegetations height (as a proxy for the complexity of the vegetations structure), and local indices of elevation variance. Using a dissimilarity approach (edge-detection, split-moving window analysis), no one variable accurately depicted both the lagg-mineral land and bog-lagg boundaries. Some indicators were better at predicting the bog-lagg boundary (i.e., vegetation height) and others at finding the lagg-mineral land boundary (i.e., topography). Dissimilarity analysis reinforces the usefulness of derived variables (e.g., wetness indices) in locating laggs, especially for those with weak topographic and vegetation gradients. When the lagg was confined between the bog and the adjacent upland, it took a linear form, parallel to the peatlands edge and was easier to predict. When the adjacent mineral land was flat or sloping away from the peatland, the lagg was discontinuous and intermittent and more difficult to predict.