Full Terms & Conditions of access and use can be found at https://www.tandfonline.com/action/journalInformation?journalCode=tres20 International Journal of Remote Sensing ISSN: 0143-1161 (Print) 1366-5901 (Online) Journal homepage: https://www.tandfonline.com/loi/tres20 Wetland classification using Radarsat-2 SAR quad- polarization and Landsat-8 OLI spectral response data: a case study in the Hudson Bay Lowlands Ecoregion Steven E. Franklin, Erik M. Skeries, Michael A. Stefanuk & Oumer S. Ahmed To cite this article: Steven E. Franklin, Erik M. Skeries, Michael A. Stefanuk & Oumer S. Ahmed (2018) Wetland classification using Radarsat-2 SAR quad-polarization and Landsat-8 OLI spectral response data: a case study in the Hudson Bay Lowlands Ecoregion, International Journal of Remote Sensing, 39:6, 1615-1627, DOI: 10.1080/01431161.2017.1410295 To link to this article: https://doi.org/10.1080/01431161.2017.1410295 Published online: 30 Nov 2017. Submit your article to this journal Article views: 331 View related articles View Crossmark data
Transcript
Full Terms & Conditions of access and use can be found athttps://www.tandfonline.com/action/journalInformation?journalCode=tres20
Wetland classification using Radarsat-2 SAR quad-polarization and Landsat-8 OLI spectral responsedata: a case study in the Hudson Bay LowlandsEcoregion
Steven E. Franklin, Erik M. Skeries, Michael A. Stefanuk & Oumer S. Ahmed
To cite this article: Steven E. Franklin, Erik M. Skeries, Michael A. Stefanuk & Oumer S. Ahmed(2018) Wetland classification using Radarsat-2 SAR quad-polarization and Landsat-8 OLI spectralresponse data: a case study in the Hudson Bay Lowlands Ecoregion, International Journal ofRemote Sensing, 39:6, 1615-1627, DOI: 10.1080/01431161.2017.1410295
To link to this article: https://doi.org/10.1080/01431161.2017.1410295
Wetland classification using Radarsat-2 SARquad-polarization and Landsat-8 OLI spectral response data:a case study in the Hudson Bay Lowlands EcoregionSteven E. Franklin a, Erik M. Skeriesa, Michael A. Stefanuka andOumer S. Ahmed a,b
aSchool of Environment, Trent University, Peterborough, Canada; bNunavik Research Center, CartographicServices, Montreal, Canada
ABSTRACTCircumboreal Canadian bogs and fens distinguished by differencesin soils, hydrology, vegetation and morphological features wereclassified using combinations of Radarsat-2 synthetic apertureradar (SAR) quad-polarization data and Landsat-8 OperationalLand Imager (OLI) spectral response patterns. Separate classifica-tions were conducted using a traditional pixel-based maximumlikelihood classifer and a machine learning algorithm following anobject-based image analysis (OBIA). This study focused on twowetland classes with extensive coverage in the area (bog andfen). In the pixel-based maximum likelihood classification, accu-racy increased from approximately 69% user’s accuracy and 79%producer’s accuracy using Radarsat-2 SAR data alone to approxi-mately 80% user’s accuracy and 87% producer’s accuracy usingLandsat-8 OLI data alone. Use of the Radarsat-2 SAR and Landsat-8OLI data following principal components analysis (PCA) datafusion did not result in higher pixel-based maximum likelihoodclassification accuracy. In the object-based machine learning clas-sification, higher bog and fen class accuracies were obtained whenusing Radarsat-2 and Landsat OLI data individually compared tothe equivalent pixel-based classification. Subsequently, a PCA-datafusion product outperformed the individual bands of the Radarsat-2 and Landsat-8 imagery in object-based classification. Greaterthan 90% producer’s accuracy was obtained. The margin of error(MOE) was less than 5% in all classifications reported here. Furtherresearch will examine alternative data fusion techniques and theaddition of Radarsat-2 SAR interferometric digital elevation model(DEM)-based geomorphometrics in object-based classification ofdifferent morphological types of bogs and fens.
ARTICLE HISTORYReceived 17 April 2017Accepted 15 November 2017
1. Introduction
Bogs and fens are important wetland types in circumboreal Canada that are recog-nized primarily by differences in soil type and constituents (e.g. fibrisols versushumisols), dominant vegetation cover (e.g. sphagnum versus sedge), nutrient status
CONTACT Steven E. Franklin, [email protected] School of Environment, Trent University, Peterborough 7B8,Canada
INTERNATIONAL JOURNAL OF REMOTE SENSING, 2018VOL. 39, NO. 6, 1615–1627https://doi.org/10.1080/01431161.2017.1410295
(e.g. ombrotrophic versus eutrophic), and slope/basin morphology (Stewart andKantrud 1971; NWWG 1997; Keddy 2010). Satellite sensor classifications of such wet-lands have improved in recent years with several methodological innovations, such asthe use of multiple sensor datasets and data fusion techniques (Grenier et al. 2007;Idol, Haack, and Mahabir 2015), implementation of object-based image analysis (OBIA)(Dingle Robertson, King, and Davies 2015), extraction of geomorphometric variablesfrom digital elevation models (DEMs) (Difebo, Richardson, and Price 2015), andapplication of more complex classifiers, such as evidential reasoning and machinelearning algorithms, rather than traditional statistical classification techniques (Millardand Richardson 2013). Further tests of these and related methodological improve-ments for remote sensing wetland classification and mapping have been recom-mended (Tiner, Lang, and Klemas 2015). In particular, additional examples of themethods and benefits of object-based machine learning classification approaches areneeded (Dronova 2015).
Typically, workflows for satellite sensor wetland object-based classification are com-plex and few guidelines exist to help users navigate the many options available depend-ing on the desired mapping scale and extent and the level of accuracy required. To citeonly one potentially important decision, an image analyst must select the sourceimagery to use in the classification; for example, satellite multispectral or syntheticaperture radar (SAR) imagery – or, increasingly – both. If a multispectral/SAR multiplesensor dataset is employed, the analyst could decide to ‘fuse’ the images to create newvariables (Jiang et al. 2011; Pohl 2016) or simply rely on the power of the classificationalgorithm to extract complementary information from individual bands of multiplesensors (Peddle et al. 1994). In addition, since wetland form and process are oftenrelated to morphometric and basin characteristics (Minar, Evans, and Krcho 2013),interferometric SAR images (or other data sources, such as lidar) may be acquired toenable DEM- based geomorphometric analysis (Short et al. 2011). The differences inresulting wetland classification accuracy based on the selection of variables alone mayor may not justify the differences in investment in software costs, training the classifier(and the personnel involved) and ease of interpretation of the results. Of course, wetlandconditions at the time of image acquisition (e.g. standing water, type, phenology,moisture content and amount of vegetation) will strongly influence classification results(Henderson and Lewis 2008; Corcoran et al. 2012).
This article introduces a wetland mapping case study in the Hudson Bay LowlandsEcoregion of northern Ontario in which Radarsat-2 SAR quad-polarization data andLandsat-8 Operational Land Imager (OLI) spectral response patterns were classified inpixel-based and object-based classification approaches in separate analyses. First, thespectral response patterns observed by the Landsat-8 OLI sensor and quad-polarizationC-band SAR data acquired by Radarsat-2 sensors were analysed in a traditional pixel-based maximum likelihood supervised classification. Second, those results were com-pared to the accuracy obtained using an object-based image analysis (OBIA) classifica-tion based on a machine learning algorithm with access to a similar set of classificatoryvariables. Results are presented for the two major wetland classes of interest – bogs andfens, which in this lowland region differ principally in their characteristic vegetation andeco-hydrological conditions (Wells and Zoltai 1985). The overall objective of this article isto interpret the classification maps and compare the classification accuracy when using
1616 S. E. FRANKLIN ET AL.
Radarsat-2 SAR and Landsat-8 satellite sensor in pixel-based maximum likelihood andobject-based machine learning classification methods.
2. Study area and data used
The study area encompasses an active open pit diamond mine in the Hudson BayLowlands in northern Ontario (Figure 1). Structural anomalies and kimberlite diamond-bearing intrusions of Archaean age occur in younger and sub-horizontally bedded
Figure 1. Location of the study area in northern Ontario, Canada.
INTERNATIONAL JOURNAL OF REMOTE SENSING 1617
Silurian limestone in this region (Morris and Kaszycki 1997). Further expansion of themine site and other resource extraction developments are planned in the near future aspart of the Ontario ‘Ring of Fire’ Mineral Development Strategy (Ontario Ministry ofNorthern Development and Mines 2015).
The area is dominated by poorly-drained organic and mineral soils, discontinuouspermafrost, shallow basal Quaternary glacial tills, and glaciolacustrine clays (CEAA 2005).The mean annual precipitation ranges between 528 and 833 mm, and the mean dailyJuly temperature ranges between 12°C and 16°C (Crins et al. 2009). The growing seasonis short, and the presence of saturated peatlands in low (but continuing positiveisostatic) topographic relief creates diverse eco-hydrological wetland features, whichcover approximately 90% of the area (Crins et al. 2009). Relatively small amounts ofstunted, upland forest are located on higher slopes and close to the shorelines of smallcreeks, riverine systems, and lentic systems (CEAA 2005). Wetlands in this region aretypically recognized as different types of fens and bogs according to the CanadianWetland Classification System (NWWG 1997). Many contain significant standing orpooled water (Keddy 2010). Fens occur in gradations of low-lying open-, patterned-and graminoid vegetation communities on saturated organic soils and peat up to 2 m indepth. Some fens have modest levels of shrub or coniferous tree cover and may beribbed or sloped (Wells and Zoltai 1985). Bogs are typically drier, with substantial build-up of peat, and are expressed as lichen-, graminoid-, shrub- and treed vegetationcommunities. Bogs occur in ombrotrophic domes, mounded or flat conditions(Peckham, Ahl, and Gower 2009; Anderson et al. 2010).
The classification variables in this study are listed in Table 1. Landsat-8 imagery wasacquired on 21 August 2014 from the United States Geological Survey (USGS) as anorthorectified and top-of-atmosphere corrected product. Radarsat-2 SAR quad-polariza-tion imagery was acquired the following day, on 22 August 2014, in fine resolutionmode with a spatial resolution of 5 m. These images were georeferenced to the pan-sharpened Landsat-8 imagery with sub-pixel root mean square error using 175 groundcontrol points per scene. Two Radarsat-2 scenes were required to cover the entire studyarea, which were georeferenced independently and mosaicked. The SAR imagery wascalibrated using a sigma-nought calibration in order to reduce both backscatter andissues stemming from altitude-calibration, and was then converted from raw units topower units (decibels) (Ulaby et al. 2014).
Additional reference data used to interpret wetland classes and develop training andvalidation areas included: (i) a Worldview-2 image acquired on 7 July 2013, (ii) a SPOT-5imagery acquired on 11 September 2006, and (iii) wetland and land cover classificationmap products for this area primarily based on earlier Landsat imagery and aerialphotography interpretations (AMEC 2004; Hogg et al. 2009; OMNRF 2014).
Table 1. Classification input variables used to classify bogs and fens in pixel-based and object-basedclassification methods.Data type Classification inputs
Classification accuracies in those mapping efforts, however, were estimated to be quitelow (e.g. 64% overall by Hogg et al. 2009).
3. Methods
3.1. Texture and data fusion
Directional invariant grey level co-occurrence matrix (GLCM) texture measures of var-iance and correlation were derived over multiple window sizes for the Radarsat-2 andLandsat-8 image data (Haralick, Shanmugam, and Dinstein 1973). Two relatively smallwindow sizes were retained for analysis (3 × 3 and 5 × 5 windows). These texturesappeared to correspond visually with the high degree of local variability in the twoclasses of interest. A number of initial data fusion tests on the original dataset usingwavelet, Brovey transformations, and other data fusion routines (Wang et al. 2005; Donget al. 2009; Jiang et al. 2011; Kumar, Sinha, and Taylor 2014; Joshi et al. 2016;Sukawattanavijit and Chen 2015; Vivone et al. 2015; Khorram et al. 2016) resulted inthe selection of the principal components analysis fusion output based on a visualinspection of the quality of the imagery. Bhattacharrya-distance statistics were calcu-lated in a sample of well-known wetland areas to support the visual interpretation; inthose areas, the PCA-data fusion showed good B-distance separability between the bogand fen wetlands. PCA-data fusion was accomplished using ENVI 5.4 (Exelis VisualSolutions 2017). Five eigenvectors explaining more than 98% of the Radarsat andLandsat image variance were retained for classification purposes.
3.2. Pixel-based maximum likelihood classification
A pixel-based supervised classification method was implemented using the originalLandsat spectral response patterns and Radarsat-2 SAR quad-polarization data in non-fusion and PCA-data fusion runs. First, individual Landsat and Radarsat image andtexture variables were classified alone and then the same training areas were used todrive a combined or PCA-data fusion classification. A total of 78 individual pixel trainingareas in the bog and fen class were sampled based on the available reference data. Theanalyst selected training areas fully contained within the wetland classes of interest (i.e.not isolated single-pixel locations). Validation of the pixel-based classification of the bogand fen classes was based on an independent sample of 81 wetlands not used in thetraining data collection following minor post-classification filtering. An illustrative exam-ple of a fen training site for the pixel-based classification is shown in Figure 2.
3.3. Object-based machine learning classification
Multiresolution image segmentation was performed using Trimble eCognitionDeveloper (Trimble 2015) with user-defined parameters of scale, shape, and compact-ness. These parameters were finalized after a number of object primitives were com-pared to the wetlands interpreted as training areas in the reference data. Differentobjects were generated by segmentation of the original imagery and data fusionproducts, which were then independently classified. As is typical in object-based
INTERNATIONAL JOURNAL OF REMOTE SENSING 1619
image analysis (e.g. Hay and Castilla 2008; He et al. 2011), the relatively subjective trial-and-error approach was based on visual interpretation of the available high resolutionimagery, existing map products, and analyst expertise in recognizing bog and fenwetland features. Overall, this approach was considered effective in determining theappropriate image segmentation parameter values to yield visually acceptable wetlandobjects for subsequent classification (see also Costa et al. 2008; Drǎguţ, Tiede, and Levick2010; Dingle Robertson, King, and Davies 2015). A class-specific multicollinearity test wasconducted to eliminate highly correlated input variables prior to executing the classifi-cation (R2 > 70).
The Random Forest machine learning classification algorithm (Breiman 2001) wasimplemented based on the R package formulation (Lantz 2013). Random Forest is a non-parametric algorithm that creates multiple decision trees for each image object and themode of the classification decision dictates the class of that object (Liaw and Wiener2002). Random Forest was selected for this case study as it has been shown to be a good‘first-choice algorithm’ because it is straightforward to train, computationally efficient,highly stable with respect to variations in model parameter values, and performs as wellas other machine learning algorithms (Cracknell and Reading 2014). As in the pixel-based method, Landsat OLI and Radarsat-2 SAR data, and then the PCA-fusion datasets,were classified separately. A total of 306 areas were used for machine learning classifiertraining, and validation was conducted with 41 independent bog and fen features.
3.4. Classification accuracy and variable importance
Confusion matrices showing overall accuracy, user’s accuracy, producer’s accuracy andmargin of error (MOE) were produced and interpreted for each classification run.Summaries of those tables for the bog and fen classes of interest are reproduced in
Figure 2. Example of a fen training area sample: (a) Landsat-8 OLI bands 5, 4, 3 displayed as falsecolour composite; (b) Worldview-2 bands 5, 3, 2 displayed as false colour composite; (c) Radarsat-2SAR HV polarization image displayed as a continuous grayscale. Image area shown approximately2.5 km2.
1620 S. E. FRANKLIN ET AL.
this article as Tables 2 and 3. While the sample sizes varied depending on the classifica-tion methods, the margin of error (MOE) was less than 5% in all classifier tests reportedhere. As others have noted, the sample sizes in classification accuracy validation tests areoften relatively small, especially in object-based classifications, which rely on homoge-neous objects rather than multiple pixel identifiers within a single feature (DingleRobertson, King, and Davies 2015). However, the validation samples in this study wereconsidered appropriate as each classification run was assessed following standardprocedures that were consistent within that classification test, i.e. in a way that wasinternally consistent with known levels of confidence (Stehman and Czaplewski 1998;Pond 2016).
4. Results
Table 2 contains the pixel-based classification accuracies obtained in this study for thebog and fen wetland classes using Radarsat-2 SAR quad-polarization data alone,Landsat-8 OLI spectral response patterns alone, and a PCA-fusion dataset of combinedRadarsat-2/Landsat-8 variables. Table 3 contains the equivalent classification resultsobtained following the object-based image analysis and machine learning classificationmethods.
The use of Radarsat-2 SAR quad-polarization data alone resulted in a pixel-basedmaximum likelihood user’s/producer’s classification accuracy for bogs and fens ofapproximately 69%/79% (Table 2). This is a good result that exceeds accuracies reportedin Canada with Radarsat-1 datasets (e.g. Grenier et al. 2007). Landsat-8 OLI spectralresponse patterns alone yielded approximately 80%/87% user’s/producer’s classificationaccuracy, a significant improvement of approximately 9% over the Radarsat-2 SARresults. This result is consistent with, or slightly better than, earlier levels of Landsatsensor wetland classification accuracy in Canada (e.g. Li and Chen 2005). The PCA-fusiondataset did not result in higher accuracies in the pixel-based maximum likelihoodclassification of the wetland classes of interest. The object-based machine learning
Table 2. Pixel-based classification user’s and producer’s classification accuracy for Radarsat-2quadpolarization data, Landsat-8 OLI spectral response patterns, and a PCA-fusion dataset basedon 78 training and 81 validation samples. Margin of error was less than 5% in all tests.Variables used Bog Fen Mean
Table 3. Object-based classification user’s and producer’s classification accuracy for Radarsat-2quadpolarization data, Landsat-8 OLI spectral response patterns, and a PCA-fusion dataset basedon 306 training and 41 validation samples. Margin of error was less than 5% in all tests.Variables used Bog Fen Mean
algorithm provided approximately 80%/84% user’s/producer’s classification accuracy forbogs and fens using Radarsat-2 SAR quad-polarization data alone (Table 3). This was asignificant improvement, more than 10% classification accuracy increase, over theRadarsat-2 SAR pixel-based maximum likelihood classification results.
The object-based machine learning classification accuracy improved when usingLandsat-8 OLI data, to approximately 87%/84% user’s/producer’s accuracy. Again, thisrepresents a significant increase over the pixel-based classification results (within anMOE of less than 5%). An additional increase of approximately 7% was observed whenthe object-based machine learning algorithm was implemented with the PCA-datafusion of Radarsat-Landsat variables (to approximately 91%/93% for user’s and produ-cer’s accuracies). These object-based classification results are similar to those obtainedrecently in other Canadian wetland mapping projects using higher spatial resolutionmultispectral data (e.g. Dingle Robertson, King, and Davies 2015).
Visual interpretation of the resulting maps (Figure 3) suggested the following generalguidelines for image analysts working on wetland remote sensing projects similar to thiscase study:
Figure 3. Final classification map output following object-based image analysis and Random Forestclassification: (a) Radarsat-2 SAR quad-polarization data alone; (b) Landsat-8 OLI data alone; (c) PCA-data fusion of Radarsat-2 SAR and Landsat-8 OLI data. Map area shown approximately 105 km2.
1622 S. E. FRANKLIN ET AL.
(1) The use of multiple sensor data, in this case the combined Radarsat-2/Landsat-8PCA-data fusion products, was preferred over the individual Radarsat and Landsatimage data when interpreting wetland extent and type.
(2) Multispectral Landsat images were frequently more readily interpreted andappeared to display wetlands in ways that were consistent, ‘realistic’ and familiarto the analysts; for example, after delineating training features in the Landsatimagery, it was less difficult to identify the same feature in the Radarsat imagery.
(3) The Radarsat-2 SAR quad-polarization classifications, while less accurate in astatistical sense, did appear to show distinctive wetland features and spatialcontext well that may be related to local wetland form (e.g. morphologicalidentifiers).
(4) A reduction in wetland feature fragmentation was noticed when object-basedimage analysis methods were employed; note that this comment is consistentwith the recent literature comparing higher spatial resolution multispectral pixel-based and object-based wetland classifications (e.g. Dingle Robertson, King, andDavies 2015)); and, finally.
(5) Object-based classifications were more accurate based on statistical tests, and alsoappeared to have fewer isolated ‘salt-and-pepper’ wetland pixels and greateroverall coherence and wetland integrity compared to the pixel-based classifica-tion maps.
Overall, the pixel-based and object-based classification results in the Hudson BayLowland Region were consistent with those reported in the literature in other wetlandmapping studies that have compared pixel-based maximum likelihood classification andobject-based classification approaches with multispectral and SAR images. The increasesin object-based classification accuracy observed in the present study based on PCA-datafusion of Radarsat and Landsat data sets also confirm the value of multiple sensors forwetland mapping as has been noted elsewhere (e.g. Rodrigues and Souza-Filho 2011). Inaddition, accuracies greater than 90% (such as obtained in this study for bogs and fens)are comparable to those reported using fusion of Radarsat-2 SAR quad-polarization dataand higher spatial resolution optical data sets, and to those which employed detailedDEM analysis to reveal wetland geomorphological and topographic characteristics (e.g.Mui, He, and Weng 2015). Additional research is now warranted to examine morepowerful data fusion techniques and to incorporate wetland geomorphological condi-tions captured in DEMs (Lecours et al. 2017). Such wetland geomorphometric analysismay be feasible through interferometric processing of the Radarsat-2 SAR data (Shortet al. 2011).
5. Conclusion
Extensive circumboreal bogs and fens in the Hudson Bay Lowlands Ecoregion of north-ern Ontario differ in eco-hydrological characteristics and dynamics and were classifiedusing Radarsat-2 SAR quad-polarization data and Landsat-8 OLI images. Pixel-basedmaximum likelihood classification and object-based machine learning classificationmethods generated good results and map output. The best accuracies were approxi-mately 85% and 93% for pixel- and object-based classifiers, respectively, with a MOE of
INTERNATIONAL JOURNAL OF REMOTE SENSING 1623
less than 5%. Significant increases were not observed when using a Radarsat-LandsatPCA data fusion product with the pixel-based maximum likelihood algorithm. Object-based machine learning classification accuracies did benefit following PCA-data fusion ofthe two datasets. Additional research in data fusion and interferometric SAR analysis ofwetlands is warranted to explore the performance of Radarsat-2 and Landsat-8 datasetsin wetland classification.
Acknowledgements
The authors are grateful to the reviewers and editors for helpful comments that improved themanuscript.
Disclosure statement
No potential conflict of interest was reported by the authors.
Funding
This research was supported by a SOAR grant from the Canadian Space Agency and a DiscoveryGrant provided by the Natural Sciences and Engineering Research Council of Canada.
ORCID
Steven E. Franklin http://orcid.org/0000-0003-1886-9153Oumer S. Ahmed http://orcid.org/0000-0001-7049-9829
References
AMEC. 2004. Victor Diamond Project Environmental Assessment Comprehensive Study (CSEA).Mississauga, ON: De Beers Canada Exploration Inc. and AMEC Consulting .
Anderson, K., J. Bennie, E. J. Milton, P. Hughes, R. Lindsay, and R. Meade. 2010. “Combining LiDARand IKONOS Data for Eco-Hydrological Classification of an Ombrotrophic Peatland.” Journal ofEnvironmental Quality 39 (1): 260–273. doi:10.2134/jeq2009.0093.
Breiman, L. 2001. “Random Forests.” Machine Learning 45 (1): 5–32. doi:10.1023/A:1010933404324.CEAA. 2005. “Victor Diamond Project Comprehensive Study Report.” Accessed 2 May 2016. https://
www.ceaa-acee.gc.ca/80C30413-docs/report_e.pdfCorcoran, J., J. Knight, B. Brisco, S. Kaya, A. Cull, and K. Murnaghan. 2012. “The Integration of
Optical, Topographic, and Radar Data for Wetland Mapping in Northern Minnesota.” CanadianJournal of Remote Sensing 37 (5): 564–582. doi:10.5589/m11-067.
Costa, G., R. Feitosa, T. Cazes, and B. Feijó. 2008. “Genetic Adaptation of SegmentationParameters.” In Object-Based Image Analysis, edited by T. Blaschke, S. Lang, and G. Hay, 679–695. BerlIn Springer.
Cracknell, M. J., and A. M. Reading. 2014. “Geological Mapping Using Remote Sensing Data: AComparison of Five Machine Learning Algorithms, Their Response to Variations in the SpatialDistribution of Training Data and the Use of Explicit Spatial Information.” Computers andGeosciences 63: 22–33. doi:10.1016/j.cageo.2013.10.008.
Crins, W. J., P. A. Gray, P. Uhlig, and M. C. Wester. 2009. The Ecosystems of Ontario, Part 1: Ecozonesand Ecoregions. Peterborough, ON: Ministry of Natural Resources, Science and InformationBranch, Inventory, Monitoring and Assessment Section Technical Report, SIB TER IMA TR-01.
Difebo, A., M. Richardson, and J. Price. 2015. “Fusion of Multispectral Imagery and LiDAR DigitalTerrain Derivatives for Ecosystem Mapping and Morphological Characterization of a NorthernPeatland Complex.” In Remote Sensing of Wetlands: Applications and Advances, edited by R. W.Tiner, M. W. Lang, and V. V. Klemas, 399–412. Boca Raton: CRC Press.
Dingle Robertson, L., D. J. King, and C. Davies. 2015. “Object-Based Image Analysis of Optical andRadar Variables for Wetland Evaluation.” International Journal of Remote Sensing 36 (23): 5811–5841. doi:10.1080/01431161.2015.1109727.
Dong, J., D. Zhuang, Y. Huang, and J. Fu. 2009. “Advances in Multi-Sensor Data Fusion: Algorithmsand Applications.” Sensors 9 (10): 7771–7784. doi:10.3390/s91007771.
Drǎguţ, L., D. Tiede, and S. R. Levick. 2010. ““ESP: A Tool to Estimate Scale Parameter forMultiresolution Image Segmentation of Remotely Sensed Data.” International Journal ofGeographical Information Science 24 (6): 859–871. doi:10.1080/13658810903174803.
Dronova, I. 2015. “Object-Based Image Analysis in Wetland Research: A Review.” Remote Sensing 7:6380–6413. doi:10.3390/rs70506380.
Grenier, M., A. Demers, S. Labrecque, M. Benoit, R. A. Fournier, and B. Drolet. 2007. “An Object-Based Method to Map Wetland Using RADARSAT-1 and Landsat ETM Images: Test Case on TwoSites in Quebec, Canada.” Canadian Journal of Remote Sensing 33 (S1): S28–S45. doi:10.5589/m07-048.
Haralick, R. M., K. Shanmugam, and I. H. Dinstein. 1973. “Textural Features for Image Classification.”IEEE Transactions on System, Man, and Cybernetics SMC-3 (6): 610–621. doi:10.1109/TSMC.1973.4309314.
Hay, G. J., and G. Castilla. 2008. “Geographic Object-Based Image Analysis (GEOBIA): A New Namefor A New Discipline.” In Object-Based Image Analysis, edited by T. Blaschke, S. Lang, and G. Hay,75–89. BerlIn Springer.
He, Y., S. E. Franklin, X. Guo, and G. B. Stenhouse. 2011. “Object-Oriented Classification of Multi-Resolution Images for the Extraction of Narrow Linear Forest Disturbance.” Remote SensingLetters 2 (2): 147–155. doi:10.1080/01431161.2010.504755.
Henderson, F. M., and A. J. Lewis. 2008. “Radar Detection of Wetland Ecosystems: A Review.”International Journal of Remote Sensing 29 (20): 5809–5835. doi:10.1080/01431160801958405.
Hogg, A., I. Smyth, G. Naoumov, N. Bookey, and P. Uhlig. 2009. Interpretive Guide to the ProvincialLandcover Classification 2000: Focus on the Far North. Peterborough, ON: Ontario Ministry ofNatural Resources, Science and Information Branch, Inventory, Monitoring and AssessmentSection.
Idol, T., B. Haack, and R. Mahabir. 2015. “Radar and Optical Remote Sensing Data Evaluation andFusion; a Case Study for Washington, DC, USA.” International Journal of Image and Data Fusion 6(2): 138–154. doi:10.1080/19479832.2015.1017541.
Jiang, D., D. Zhuang, J. Fu, and Y. Huang. 2011. “A Survey of Multispectral Image FusionTechniques in Remote Sensing Applications.” In Image Fusion and Its Applications, edited byY. Zheng, 1–22. India: InTech.
Joshi, N., M. Baumann, A. Ehammer, R. Fensholt, K. Grogan, P. Hostert, M. Rudbeck Jepsen, et al.2016. “A Review of the Application of Optical and Radar Remote Sensing Data Fusion to LandUse Mapping and Monitoring.” Remote Sensing 8 (1): 60–70. doi:10.3390/rs8010070.
Keddy, P. A. 2010. Wetland Ecology: Principles and Conservation. Chicago: Cambridge UniversityPress.
Khorram, S., C. F. van der Wiele, F. H. Koch, S. A. Nelson, and M. D. Potts. 2016. Principles of AppliedRemote Sensing. New York: Springer.
Kumar, L., P. Sinha, and S. Taylor. 2014. “Improving Image Classification in a Complex WetlandEcosystem through Image Fusion Techniques.” Journal of Applied Remote Sensing 8 (1).doi:10.2227/1.JRS.8.083616.
Lantz, B. 2013. “Machine Learning with R.” Birmingham, UK: Open Source: Packt.Lecours, V., R. Devillers, A. E. Simms, V. Lucieer, and C. Brown. 2017. “Toward a Framework for
Terrain Attribute Selection in Environmental Studies.” Environmental Modelling and Software 89:19–30. doi:10.1016/j.envsoft.2016.11.027.
Li, J., and W. Chen. 2005. “A Rule-Based Method for Mapping Canada’s Wetlands Using Optical,Radar and DEM Data.” International Journal of Remote Sensing 26 (22): 5051–5069. doi:10.1080/01431160500166516.
Liaw, A., and M. Wiener. 2002. “Classification and Regression by Random Forest.” R News 2 (3):18–22.
Millard, K., and M. Richardson. 2013. “Wetland Mapping with LiDAR Derivatives, SAR PolarimetricDecompositions, and LiDAR–SAR Fusion Using a Random Forest Classifier.” Canadian Journal ofRemote Sensing 39 (4): 290–307. doi:10.5589/m13-038.
Minar, J., I. S. Evans, and J. Krcho. 2013. “Geomorphometry: Quantitative Land Surface Analysis.” InTreatise on Geomorphology: Methods in Geomorphology, edited by John F. Shroder, 22–34. Vol.14. San Diego: Academic Press.
Morris, T. F., and C. Kaszycki. 1997. Prospector’s Guide to Drift Prospecting for Diamonds, NorthernOntario. Toronto: Ontario Geological Survey, Miscellaneous Paper 167.
Mui, A., Y. He, and Q. Weng. 2015. “An Object-Based Approach to Delineate Wetlands acrossLandscapes of Varied Disturbance with High Spatial Resolution Satellite Imagery.” Journal ofPhotogrammetry and Remote Sensing 109: 30–46. doi:10.1016/j.isprsjprs.2015.08.005.
NWWG (National Wetlands Working Group). 1997. The Canadian Wetland Classification System:. 2nded. Waterloo: Wetlands Research Centre, University of Waterloo.
OMNRF (Ontario Ministry of Natural Resources and Forestry). 2014. Far North Land Cover V1.4(Dataset). Toronto, ON: Queen's Printer for Ontario.
Ontario Ministry of Northern Development and Mines. 2015. Ontario’s Mineral DevelopmentStrategy. Sudbury, ON: Queen’s Printer for Ontario.
Peckham, S. D., D. E. Ahl, and S. T. Gower. 2009. “Bryophyte Cover Estimation in a Boreal BlackSpruce Forest Using Airborne Lidar and Multispectral Sensors.” Remote Sensing of Environment113 (6): 1127–1132. doi:10.1016/j.rse.2009.02.008.
Peddle, D. R., G. M. Foody, A. Zhang, S. E. Franklin, and E. F. LeDrew. 1994. “Multi-Source ImageClassification II: An Empirical Comparison of Evidential Reasoning and Neural NetworkApproaches.” Canadian Journal of Remote Sensing 20 (4): 396–407. doi:10.1080/07038992.1994.10874582.
Pohl, C. 2016. “Multisensor Data Fusion Guidelines in Remote Sensing”. 9th InternationalSymposium of the International Society for Digital Earth. IOP Conf. Series: Earth andEnvironmental Science. doi:10.1088/1755-1315/34/1/012026.
Pond, B. 2016. “Across the Grain: Multi-Scale Map Comparison and Land Change Assessment.”Ecological Indicators 71: 660–668. doi:10.1016/j.ecolind.2016.06.056.
Rodrigues, S. W. P., and P. W. Souza-Filho. 2011. “Use of Multi-Sensor Data to Identify and MapTropical Coastal Wetlands in the Amazon of Northern Brazil.” Wetlands 31 (1): 11–23.doi:10.1007/s13157-010-0135-6.
Short, N., B. Brisco, N. Couture, W. Pollard, K. Murnaghan, and P. Budkewitsch. 2011. “A Comparisonof Terra-SAR-X, Radarsat-2 and ALOS-PALSAR Interferometry for Monitoring PermafrostEnvironments: A Case Study from Herschel Island, Canada.” Remote Sensing of Environment115: 3491–3506. doi:10.1016/j.rse.2011.08.012.
Stehman, S., and R. Czaplewski. 1998. “Design and Analysis for Thematic Map AccuracyAssessment: Fundamental Principles.” Remote Sensing of Environment 64: 331–344.doi:10.1016/S0034-4257(98)00010-8.
Stewart, R. E., and H. A. Kantrud. 1971. Classification of Natural Ponds and Lakes in the GlaciatedPrairie Region. Washington, DC: USDA Resource Publ. 92, Bureau of Sport Fishers and Wildlife.
Sukawattanavijit, C., and J. Chen. 2015. “Fusion of RADARSAT-2 Imagery with LANDSAT-8Multispectral Data for Improving Land Cover Classification Performance Using SVM.” InProceedings, IEEE 5th Asia-Pacific Conference on Synthetic Aperture Radar (APSAR), edited by K.Y. See, 567–572. Singapore: IEEE, Research Publishing Services.
Tiner, R. W., M. W. Lang, and V. V. Klemas. 2015. Remote Sensing of Wetlands: Applications andAdvances. Boca Raton: CRC Press.
Ulaby, F. T., D. G. Long, W. J. Blackwell, C. Elachi, A. K. Fung, C. Ruf, K. Sarabandi, H. A. Zebker, and J.Van Zyl. 2014. Microwave Radar and Radiometric Remote Sensing. London: Artech House.
Vivone, G., L. Alparone, J. Chanussot, M. Dalla Mura, A. Garzelli, G. A. Licciardi, R. Restaino, and L.Wald. 2015. “A Critical Comparison among Pansharpening Algorithms.” IEEE Transactions onGeoscience and Remote Sensing 53 (5): 2565–2586. doi:10.1109/TGRS.2014.2361734.
Wang, Z., D. Ziou, C. Armenakis, D. Li, and Q. Li. 2005. “A Comparative Analysis of Image FusionMethods.” IEEE Transactions on Geoscience and Remote Sensing 43 (6): 1391–1402. doi:10.1109/TGRS.2005.846874.
Wells, E. D., and S. Zoltai. 1985. “The Canadian System of Wetland Classification and Its Applicationto Circumboreal Wetlands.” Aquilo Series Botanic 21: 45–52.
