Recent and Emerging Methodological Approaches to High-Latitude Vegetation Mapping: An Annotated Bibliography
Compiled in Fall 2025 by Bonnie Bernard for
GEOG 560: Introduction to Geographic Information Science
Photo credit: Bonnie Bernard, all images.
Review Papers
1. Stow, D. A., Hope, A., McGuire, D., Verbyla, D., Gamon, J., Huemmrich, F., … & Myneni, R. (2004). Remote sensing of vegetation and land-cover change in Arctic Tundra Ecosystems. Remote sensing of environment, 89(3), 281-308. [https://doi.org/10.1016/j.rse.2003.10.018]
Summary:
This review paper provided an excellent overview of the state of remote sensing approaches to arctic vegetation mapping at the beginning of the 21st century. It opened with a summary of concepts essential to landscape analyses of high-latitude ecosystems, including fundamental principles of spectral change detection and spectral indices, as well as the unique challenges associated with remote imaging of the Arctic. The authors first devoted a section to reviewing the then-best available satellite data, as well as spectral indices useful for indexing plant biophysical parameters like biomass. In the second section, they summarized imaging complications unique the Arctic, such as its short growing seasons, frequent cloud and snow cover, dynamic sun angles, and abundance of small shallow lakes and seasonal waters.
Following this concept review, the authors examined nine case studies that used multi-temporal imagery to quantify Arctic ecosystem change via a range of methodologies, imagery resolutions, and spatial and temporal extents. At the time, imagery provided by the Advanced Very High Resolution Radiometer (AVHRR) was most commonly used, although Moderate Resolution Imaging Spectroradiometer (MODIS) imagery was beginning to gain popularity. A few of the case studies they reviewed also included analyses based on ground-level photography or repeat aerial photography. These case studies provided early insight into seasonal NDVI dynamics, landcover mapping, shrub cover expansion, permafrost thaw, and impacts of landscape change on wildlife populations.
Topical Relevance and Methodological Insights:
In general, I sought to use review papers as a means of getting quickly up-to-speed on the concepts and foundational studies published relevant to my topic before ~2015, and as a means of identifying keystone papers worth reading separately. This review provided an excellent framework for understanding the unique challenges of obtaining quality optical data in high-latitude ecosystems, and it provided me with essential background for contextualizing more recent methodological advances in mapping methodologies. The case studies highlighted within also provided a good summary of the ecological motivations— namely shrub expansion and permafrost thaw— for many of the developments that have occurred in recent years.
2. Xie, Y., Sha, Z. and Yu, M., 2008. Remote sensing imagery in vegetation mapping: a review. Journal of plant ecology, 1(1), pp.9-23. [https://doi.org/10.1093/jpe/rtm005]
Summary:
In this review, Xie et al. provided an overview of widely used remote sensing workflows and methods for mapping vegetation cover. Specifically, they compared common imagery sources, reviewed data processing techniques, and examined common modeling and accuracy assessment methodologies. They also reviewed the general strengths and limitations of using remote sensing to map vegetation mapping, and emphasized the importance of understanding the appropriate use case for each dataset and methodology.
Topical Relevance and Methodological Insights:
This review was an excellent refresher on/ introduction to the foundations of vegetation mapping theory and methodology. Of particular usefulness to me were (1) the descriptions of older remote sensing sensors like AVHRR and MODIS, from which many foundational maps and methodologies were developed; (2) the overview of image preprocessing techniques and the extent to which different agencies provided “analysis-ready” imagery; and (3) the review of traditional statistical approaches to image classification as a baseline for learning about more recent improvements to classification and mapping methodology.
3. Ustin, S. L., & Gamon, J. A. (2010). Remote sensing of plant functional types. New Phytologist, 186(4), 795-816. [https://doi.org/10.1111/j.1469-8137.2010.03284.x]
Summary:
Ustin and Gamon (2010) provided a thorough review of the plant functional type (PFT) concept, from both ecological and remote sensing perspectives, and argued in favor of PFTs over categorical mapping. To this end, they first emphasized how PFTs preserve information about ecologically important vegetation characteristics, while still reducing complexity enough to facilitate landscape-scale analyses. The authors then reviewed the general principles of the ‘functional convergence hypothesis’— the idea that plant structural traits reflect resource availability, such that vegetation traits detectable by remote sensing can be meaningfully related to physiological traits characteristic of functional types. For example, there is a consistent, measurable relationship between absorbed photosynthetically active radiation (APAR), which can be measured optically, and net primary production (NPP), which is a measure of plant biomass acquisition. The authors noted that dominant overstory species tend to exert outsized influence on ecosystem processes, and concluded that this represents a synergy with most spectral remote sensing approaches, which primarily capture upper-canopy reflectance.
Ustin and Gamon also provided a brief review of the history of remote sensing-based vegetation mapping, beginning with the color-infrared photography based approach of the mid-20th century. They recounted how computer-based image-analysis approaches arose following contemporaneous developments in informatics and digital sensors— specifically, how digital photography increased spectral range and how developments in computation facilitated automation of processing complex data. As early as 1995, other investigators (e.g. DeFries et al. 1995, reviewed below) had begun developing continuous field approaches that would pave the way for continuous foliar cover and high-resolution PFT mapping. These methods, broadly termed “mixture analysis methods,” were developed to address the mixed-pixel problem— i.e., the reality that more than one vegetation type often occurs within a single pixel, which in turn creates complications for classification.
Finally, the authors argued for the adoption of a more integrated term for PFTs in the era of remote sensing: ‘optical type.’ To Ustin and Gamon, ‘optical types’ were plant groupings based on spectral properties with demonstrable ties to ecologically important plant physiological phenomena. While the term itself seems to have been only occasionally adopted, the concept has remained essential in the field of continuous cover mapping.
Topical Relevance and Methodological Insights:
I found this review to be an extremely thorough-yet-approachable overview of the plant functional type concept, and I appreciated their explanation of its ties to spectral indices. It was particularly valuable to me in understanding why certain indices are standard for certain ecological applications, and helped me to wrap my head around some of the newer indices.
I would be interested to read more theoretical work on the relationship between ecosystem influence and canopy dominance that the authors accept as a premise. Given that any-hit cover estimation methods can perform similarly to or better than top-cover methods for continuous field mapping (Karl et al. 2017, reviewed below), I suspect the relationship is significantly more complicated than is summarized here. I would expect those kinds of subtleties to be particularly important for high-latitude ecosystems for questions related to shrub encroachment.
4. Coops, N. C., & Wulder, M. A. (2019). Breaking the habit (at). Trends in Ecology & Evolution, 34(7), 585-587. [https://doi.org/10.1016/j.tree.2019.04.013]
Summary:
In this review, Coops and Wulder advocated for a field-wide movement away from categorical vegetation and habitat maps toward a continuous approach to mapping, with an emphasis on modeling landscape change. They acknowledged traditional criticisms of categorical vegetation mapping (e.g., difficulties reaching consensus when defining classes, incorporating categorical data into traditional statistical approaches, etc) but they focused on the power of continuous mapping to drive data-driven hypotheses about ecosystem change across space and time. In essence, their argument is one of resolution: when aggregating vegetation maps into discrete classes, practitioners (1) inherently reduce the amount of information conveyed across the coverage area, and (2) also reduce the power and number of statistical tools available for change detection.
Coops and Wulder emphasized the value of continuous field mapping for change detection by highlighting that substantial variation can occur within a categorical map class, and that vegetation shifts are likely to occur within that range of variability, rather than as a transition from one class to another. Such an approach allows exploration of dynamic vegetation responses after disturbances such as fire or pest outbreaks. Finally, map products derived from surface reflectance data, they note, also allow for the calculation of error estimates when conducting change analyses.
Topical Relevance and Methodological Insights:
For me, this paper really motivated the analytical and statistical basis for the movement toward fractional land cover and continuous field mapping approaches. Of particular interest to me was the argument that by the time the vegetation encompassed by a pixel has changed enough to transition from one map category to another, numerous more subtle shifts in composition have likely occurred and been missed.
Primary Literature
5. DeFries, R. S., Field, C. B., Fung, I., Justice, C. O., Los, S., Matson, P. A., … & Vitousek, P. M. (1995). Mapping the land surface for global atmosphere‐biosphere models: Toward continuous distributions of vegetation’s functional properties. Journal of Geophysical Research: Atmospheres, 100(D10), 20867-20882. [https://doi.org/10.1029/95JD01536]
Summary
This present work by DeFries et al. (1995) serves as a keystone paper in the literature on continuous-vegetation cover mapping and represents one of the earliest high-impact arguments for moving away from discrete categorical mapping. In this paper DeFries et al. reviewed of a number of then-common land surface classification schemes and discussed their shortcomings and discontinuities. To demonstrate how important spatial variability in ecological parameters is lost through categorical mapping, the authors then engaged in a series of analytic exercises in which they compared continuous-field estimates of ecological parameters with estimates reconstructed from categorical maps. For example, they calculated global NDVI data as a continuous field and compared the distribution to NDVI values calculated as the weighted-sum estimate of mean NDVI by vegetation class. DeFries et al. similarly compared the distributions of projected biomass, leaf area index, and albedo estimated from categorical vegetation maps to the values predicted by their spectrally-derived (NDVI-based) map. By graphing the distribution of estimates produced via contrasting methods across multiple parameters, DeFries et al. clearly demonstrated the loss of ecologically relevant information that occurs when complex vegetation metrics are collapsed into discrete classes.
The second central question of this paper was whether remote sensing tools could provide estimates of water, energy, and/or carbon exchange adequate to replace the categorical proxy of vegetation class. To this end, the authors assessed then-present methods of measuring absorbed radiation, energy and water exchange, photosynthesis, NPP, and nutrient dynamics, and discuss how each might be mapped as a continuous-field property. They argued that, taken together, these “continuous distributions of a small number of functional properties” provided a better parameterization of these metrics for modeling. Finally, DeFries et al. proposed a suite of vegetation characteristics that might be measured and georeferenced to build continuous-field maps of these biophysical properties– namely, woodiness, seasonality, leaf type, photosynthetic pathway, plant longevity, and disturbance type. As illustrated in later-reviewed papers, these characteristics would go on to form the basis of numerous plant functional type (PFT) groupings.
Topical Relevance and Methodological Insights:
Though the methods in this paper were relatively simple analytical exercises, the visionary nature of this paper really stood out to me. Even thirty years ago, these researchers at the forefront of biogeosciences were beginning to recognize the revolutionary potential of remote sensing to reshape landscape ecology and global change analysis. That the authors saw applications for remote sensing technologies before they yet existed, some of which are only just being developed in earnest, is extremely impressive to me. In some ways, their argument is essentially one urging ecologists to embrace and make full use of the field data model, which the remote sensing data they are using already follows.
6. DeFries, R. S., Townshend, J. R. G., & Hansen, M. C. (1999). Continuous fields of vegetation characteristics at the global scale at 1‐km resolution. Journal of Geophysical Research: Atmospheres, 104(D14), 16911-16923. [https://doi.org/10.1029/1999JD900057]
Summary
Following up on previous work in this lab, DeFries et al. (1999) used 1-km AHVRR data to generate a global-coverage map of continuous landcover-related parameters outlined in DeFries et al. (1995)– specifically, those related to life form, leaf type, and leaf duration. This work provided one of the earliest demonstrations of spectral unmixing as a method of deriving proportional cover via spectral characteristics, and the authors intended it to serve as a prototype for the then-future MODIS mission. In total, they mapped three continuous fields related to growth form– specifically, proportion of area occupied by woody vegetation, herbaceous vegetation, and bare ground– and four continuous fields related to leaf physiology, for woody vegetation– specifically, proportion of woody vegetation that is needleleaf or broadleaf, and proportion of woody vegetation that is evergreen or deciduous. In the final map, each pixel thus had a value for each of these seven characteristics. The authors derived training data– i.e., spectrally “pure” end-member pixels– for this exercise from higher-resolution Landsat MSS imagery.
Topical Relevance and Methodological Insights:
At the time of this paper’s writing, there wasn’t solid consensus on how to quantitatively validate this kind of modeling approach, given both the newness of the methodology and the absence of global-coverage data sets on proportional cover. The approach that DeFries et al. used here was to compare their results with regional land cover classifications and compare the general geographic distributions of dominant vegetation types; since then, this approach has become standard practice in the absence of field-based verification data. One facet was of particular relevance to my interest in high-latitude vegetation mapping: boreal forest regions presented the only notable discrepancy, whereby the proportion of woody vegetation field was consistently lower than in other map products. The authors interpret this outcome as a result of the presence of numerous small lakes throughout the taiga, which when rendered at 1km resolution, decreased the overall woody coverage estimates. This was an interesting and drastic example of a mixed-pixel problem.
7. Fernandes, R., Fraser, R., Latifovic, R., Cihlar, J., Beaubien, J., & Du, Y. (2004). Approaches to fractional land cover and continuous field mapping: A comparative assessment over the BOREAS study region. Remote Sensing of Environment, 89(2), 234-251. [https://doi.org/10.1016/j.rse.2002.06.006]
Summary:
Fernandes et al. (2004) built on the continuous field approach of DeFries et al. (1995, 1999), but in a specifically high-latitude context. This paper evaluated the performance of five algorithms for subpixel mixture estimation, including both fractional cover mapping and continuous field mapping. Specifically, the authors sought to (1) determine which algorithm provided the most accurate estimate of subpixel cover and (2) compare mapping performance when using nearby (< 100km) versus distant (> 400km) training and validation data. Upon evaluating model accuracy against existing maps generated from higher-resolution imagery, Fernandes et al. (2004) found that the traditional “hard” classification approach performed worst, regardless of the proximity of training data, and note that classification produced substantial bias in its prediction of both fractional cover and continuous fields. Of the algorithmic methods they tested, they found multivariate regression and artificial neural network models to be similarly high performers when using nearby training data, and linear least squares inversion to perform best for distant data sets.
Topical Relevance and Methodological Insights:
This paper was the first to really drive home the terminology differences between fractional cover mapping and continuous field mapping for me. With fractional cover mapping, some non-overlapping binary variable (e.g., forested v non-forested) is summarized at the pixel scale for land cover classification; with continuous field mapping, multiple biophysical parameter(s) are displayed as continuous variables that can overlap within a given pixel to describe vegetaion properties. This study also directly addressed the mixed-pixel related issues with small water bodies encountered by DeFries et al. (1999), and it was among the first of those I read to highlight the impact that proximity of training & validation data (or reference pixels) can have on model performance and outcomes.
8. Olthof, I., & Fraser, R. H. (2007). Mapping northern land cover fractions using Landsat ETM+. Remote Sensing of Environment, 107(3), 496-509. [https://doi.org/10.1016/j.rse.2006.10.009]
Summary:
Olthof and Fraser (2007) here provided early methodological work on fractional cover estimation in a high-latitude ecosystem, with the goal of evaluating the reliability of arctic map products obtained from 90m Landsat data for use in biophysical modeling. This paper compared four methods for deriving land cover fractions of five PFTs (bare ground, graminoid, deciduous shrub, conifer, and water), and conducted error analyses on each method. Specifically, Olthof and Fraser compared linear least squares inversion (LLSI), linear regression, and regression trees. Within each modeling approach, they also varied the following: number of spectral bands (3 versus 5), imagery resolution (30m versus 90m), and source of calibration/validation data (local versus global).
For all modeling methods, Olthof and Fraser used Ikonos imagery at 1m spatial resolution, and a three-band, multi-temporal Landsat mosaic at 90m spatial resolution. Once modeled, they evaluated fractional cover against higher resolution imagery and field data at three sites along a 1100km north-south gradient. In general, the linear regression model with five spectral bands slightly outperformed the five-band regression tree model, which slightly outperformed the three-band regression tree and linear regression models. The linear least squares inversion (LLSI) approach performed notably worse than the other modelling approaches by all three metrics (RMSE, R2, and bias), and thus was not used in their subsequent mapping analysis. With regard to fractional mapping, regression tree models performed best, regardless of imagery resolution, except for the shrub PFT. Thus, the authors used a combined model to generate their final maps: a linear regression model for the shrub cover type, and regression tree models for the other PFTs.
Topical Relevance and Methodological Insights:
Among the papers I reviewed, Olthof and Fraser are the first to detail and address limitations related to traditional spectral unmixing– namely, (1) that LLSI is mathematically limited in the number of end-members it can discern by the the number of spectral bands in the input image; and (2) the assumption that end-member reference spectra are representative across all parts of the imagery, when in reality end-member signatures tend to be spatially dependent. Also, previous work by Olthof demonstrated that, in northern Canada, this spatial dependence is directional. The authors thus specifically sought ways to address this problem of end-member spatial dependence, and evaluated multiple linear regression and regression tree methods as an alternative to LLSI.
I found this to be an interesting and essential extension of the previous work Fernandes et al. (2004) examining the effect of training/ validation data proximity. Interestingly, Fernandes et al. (2004) and Olthof and Fraser (2007) have critically different operational definitions of “local” for Fernandes et al. (2004), nearby sites were those within 100km and distant sites, beyond; but for Olthof and Fraser (2007), ‘local’ test/ validation data were those collected at the same geographic location as the modeled pixel. These methodological issues regarding adequately representative training and validation data are essential for coherent mapping of high-latitude ecosystems, as field data are especially difficult to obtain for such challenging and remote locations.
9. Raynolds, M. K., Walker, D. A., Epstein, H. E., Pinzon, J. E., & Tucker, C. J. (2012). A new estimate of tundra-biome phytomass from trans-Arctic field data and AVHRR NDVI. Remote sensing letters, 3(5), 403-411. [https://doi.org/10.1080/01431161.2011.609188]
Summary:
This study provided essential evidence in support of a meaningful link in Arctic ecosystems between the normalized difference vegetation index (NDVI)— which had long been interpreted as a measure of photosynthetic capacity and used as a proxy for plant biomass— and actual field measurements of aboveground biomass. To this end, Raynolds et al. first reprocessed and recalibrated data from Advanced Very High Resolution Radiometer (AVHRR)8-km imagery to compile the first standardized NDVI data set with Arctic coverage. The authors then established two very large transects (> 1500km) across Arctic regions in Eurasia and North America, along which they sampled aboveground biomass from five representative quadrats at 13 sites representative of each bioclimactic subzone present. At each sampled site, the authors extracted NDVI values from two datasets: the maximum annual NDVI value for the year of vegetation sampling as derived from the reprocessed 8-km AHVRR dataset, and the maximum growing season NDVI value from 1993-1995 as derived from 1-km AHVRR imagery. Finally, they assessed the relationship between site-level biomass measurements and NDVI estimates via logarithmic regression, and then used the regression model to map projected biomass across all pixels in the corresponding images.
For both NDVI datasets, biomass and NDVI exhibited a strong (R2 > 0.9) and highly significant correlation (p < 0.001). When used to model phytomass using 2010 imagery, these models facilitated mapping biomass by bioclimactic subzone across the entirety of the arctic. In total, these models estimated overall arctic plant biomass at approximately 2.024 × 10^12 kg, with most of this biomass in Russia (46%) or Canada (38%).
Topical Relevance and Methodological Insights:
I found it highly surprising that this was the first study to quantitatively validate a consistent, significant relationship between NDVI and biomass for the Arctic, and I also found the strength of the relationship surprising given the heterogeneity of vegetation in arctic tundra. For me, this study also underscored the importance of improving non-vascular mapping methods in that Raynolds et al. found bryophytes to comprise majority of the biomass at numerous sites. On a less-formal note, this paper contained maybe the most relatable methods-section sentence I’ve read in years: “No sorting was done in the field, as Arctic field conditions are not conducive to careful sorting.” I’m beginning to wonder if/when remote sensing resolutions will be sensitive enough to detect arctic mosquito clouds…
10. Huemmrich, K. F., Gamon, J. A., Tweedie, C. E., Campbell, P. K. E., Landis, D. R., & Middleton, E. M. (2013). Arctic tundra vegetation functional types based on photosynthetic physiology and optical properties. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 6(2), 265-275. [https://doi.org/10.1109/JSTARS.2013.2253446]
Summary:
In this study, Huemmrich et al. used a spectral unmixing approach to generate continuous field maps of three plant functional types– lichens, mosses, and vascular plants– in the graminoid tundra near Utqiaġvik, Alaska. They also used plot-scale field measurements of light use efficiency (LUE) and spectral reflectance to map estimated LUE across the study area. Specifically, Huemmrich et al. established 100m field transects, along which they recorded spectral reflectance and actual vegetation cover. Using these plot-scale data, the authors were able to successfully separate each PFT via spectral unmixing based on reflectance characteristics. Huemmrich et al. then applied this spectral unmixing approach to 30m remote sensing data obtained by Earth Observing-1 Hyperion to create continuous field maps of each PFT. They also used plot-level data to model LUE values for each pixel. Importantly, they found that mosses and lichens exhibited light use efficiency within one order of magnitude of vascular plants.
Topical Relevance and Methodological Insights:
This study provides an early example of both (1) continuous-field mapping of non-vascular vegetation types and (2) up-scaling from hand-held spectroscopy measurements. Their success mapping lichens and mosses is in part attributable to the structure of the graminoid tundra, which does not form canopy structures dense enough to obstruct imaging of understory vegetation. The authors acknowledge this limitation, but in the broader context of measuring ecologically important PFTs at high latitudes, it should also be noted that this approach is extensible to non-vascular mapping only in other regions where herbaceous and shrub cover are very low. To me, their measurements of significant non-vascular photosynthetic activity (LUE), however, underscore the importance of developing methods for estimating non-vascular cover and including non-vascular species in broader ecosystem modeling.
11. Karl, J. W., McCord, S. E., & Hadley, B. C. (2017). A comparison of cover calculation techniques for relating point-intercept vegetation sampling to remote sensing imagery. Ecological Indicators, 73, 156-165. [https://doi.org/10.1016/j.ecolind.2016.09.034]
Summary:
Numerous methods of estimating plant cover exist, and important variations exist even within point-intercept sampling techniques. In this study, Karl et al. (2017) sought to compare the performance of “any-hit” and “top-hit” approaches to cover estimation in the training and development of models built from remote sensing data. Any-hit methods are those that include vegetation in any stratum in calculating total cover, whereas top-hit approaches consider only the most vertically prominent vegetation hit at a given point. To this end, they used RapidEye 5m-resolution imagery to model cover and assess both approaches across six vegetation indicators (roughly equivalent to PFTs in this context), in two different study areas. The first study area was a region of northwest Colorado, while the second study area was a region in northern California.
For northwestern Colorado, they found that any-hit cover estimates were significantly higher and more variable than top-hit cover estimates for four of the six indicators measured. However, in northern California, they found minimal and/or insignificant differences between the two cover estimation methods. In general, model performance assessments indicated that any-hit cover methods performed as well as or better than top-hit methods across all indicators, though the magnitude of performance improvement varied by indicator and study area.
Topical Relevance and Methodological Insights:
I found this paper to be a fascinating read and was mildly disappointed to find it cited only a few times in more recent papers! Although the results do not indicate the clear superiority of either cover-estimation method for validating remote sensing data, the results suggest that the structure and composition of the vegetation type have substantial influence. Differences in vertical vegetation structure are particularly relevant to high-latitude studies of shrub encroachment and treeline advance, and to growing efforts to better map non-vascular and other understory vegetation types.
12. Macander, M. J., Frost, G. V., Nelson, P. R., & Swingley, C. S. (2017). Regional quantitative cover mapping of tundra plant functional types in Arctic Alaska. Remote Sensing, 9(10), 1024. [https://doi.org/10.3390/rs9101024]
Summary:
Macander et al. (2017) used a random-forest modeling approach to map nine plant functional types (PFTs) in the arctic tundra north of the Brooks Range. Specifically, they generated continuous-field cover maps across 125,000 km2 at 30m resolution using multi-temporal Landsat data, five gridded environmental predictors, and 225 validation plots. They modeled each response variable (PFT) separately, then assessed predictor importance using conditional random forest and ranked predictor variables. Overall, their random forest models performed best for canopy-forming PFTs like deciduous shrubs, as is common with spectral remote sensing approaches, which definitionally emphasize surface reflectance. However, their multi-temporal approach demonstrated promise in addressing the challenge of modeling non-vascular PFTs and other low-stature vegetation.
For all PFTs, spectral predictors from the Landsat seasonal composites consistently outperformed environmental predictors in importance rank. Specifically, the most important predictors were spectral indices– not surface reflectance data. EVI2 (Enhanced Vegetation Index-2) and NDVI (Normalized Difference Vegetation Index) were the most frequently high-ranked predictors and were particularly important for shrub PFTs, while NDSI (Normalized Difference Snow Index) was important for lichen models. Predictors from the early July Landsat composite provided most of the top predictors for the deciduous shrub PFTs cover model, whereas for dwarf shrub PFT, lichen PFT, and total non-vascular cover models, top predictor variables came from the spring snow-free composite. These results make sense ecologically: once fully leafed-out, deciduous shrubs dominate photosynthetically active spectral signatures, whereas evergreen dwarf shrubs and thawing non-vascular plants will dominate before leaf out (spring) and after senescence (early fall).
Topical Relevance and Methodological Insights:
From this paper, I noted a number of methodological choices with regard to imagery curation and compilation for high-latitude systems that had not been detailed in previous papers. Specifically, Macander et al. (2017) generated six seasonal Landsat composites by using the median reflectance value for each pixel, except in the spring composite image, where they used the 20th percentile reflectance value because snow is typically present. Because high-latitudes are notorious for snow- and cloud-related imagery obfuscation, this is a valuable detail. The authors also calculated multiple spectral indices from their Landsat composites, despite the redundancy between some indices, to improve resolution between similar plant communities–e.g., EVI2 is useful in high-biomass communities where NDVI might become saturated. They also calculated a net NDVI difference between spring and summer to help distinguish communities with high primary productivity. These latter methodological details are particularly useful given the short duration of the growing season at high latitudes, and the differences in phenology that can occur between lowland, upland, and alpine vegetation communities.
13. Berner, L. T., Jantz, P., Tape, K. D., & Goetz, S. J. (2018). Tundra plant above-ground biomass and shrub dominance mapped across the North Slope of Alaska. Environmental Research Letters, 13(3), 035002. [https://doi.org/10.1088/1748-9326/aaaa9a]
Summary:
Berner et al. (2018) used Landsat imagery and field biomass measurements to map above-ground biomass (AGB) and shrub dominance across the North Slope of Alaska at 30m resolution. Specifically, they measured shrub-only AGB and all plants ABG at 28 field sites, and then used linear regression to model both AGB response variables as a function of site-level NDVI. The authors then compiled growing season Landsat imagery from 2007-2016 into a 1721-scene NDVI mosaic, and used the AGB regression models to map plant-AGB and shrub-ABG across the 152,000 km2 study area. From each pixel’s ratio of shrub: total plant ABG, they were also able to calculate shrub dominance (%) as a continuous field variable over the mapping area. When assessing shrub AGB model performance, Berner et al. found strong correlation to field measurements of shrub canopy height and to other, regional maps of shrub cover. In examining environmental covariates, they also found that shrub dominance increased with maximum summer air temperatures, which is in line with other studies predicting increased shrub growth under warmer future climate scenarios.
Topical Relevance and Methodological Insights:
This paper provided an excellent overview of their imagery processing methodology: in creating the composite Landsat mosaic, Berner et al. first used the Function of mask (Fmask) algorithm to mask pixels flagged as water, snow, cloud, or shadow. As with the use of modified reflectance metrics during snowy seasons by Macander et al. (2017), this is useful methodological insight for working with high-latitude imagery, which is notorious for cloud- and snow-related obstruction. After masking, Berner et al. computed NDVI across all scenes and, from these, computed the 80th percentile NDVI value for each pixel. The final composite mosaic was composed of the scenes with NDVI values closest to the computed 80th percentile value for each pixel.
14. Nawrocki, T. W., Carlson, M. L., Osnas, J. L., Trammell, E. J., & Witmer, F. D. (2020). Regional mapping of species‐level continuous foliar cover: beyond categorical vegetation mapping. Ecological Applications, 30(4), e02081. [https://doi.org/10.1002/eap.2081]
Summary:
In line with the above call by Coops and Wulder (2019), Nawrocki et al. (2020) here extend the continuous field mapping approach to model foliar cover at the species level, rather than aggregating species into plant functional types. In motivating the argument for this approach, the authors acknowledge that continuous PFT mapping approaches are useful for modeling ecosystem processes like carbon sequestration or fuel loads, but are inadequate for exploring relationships between species and environmental gradients– i.e., those related to realized niche. A “niche-based gradient approach to vegetation mapping,” they argue, facilitates questions about inter-species interactions, plant-soil responses, and relationships to wildlife.
To this end, Nawrocki et al. used a Bayesian statistical learning approach to model the relationship between field-derived measurements of six species’ foliar cover and 24 covariates, including 6 Landsat-derived spectral indices (NDVI, EVI2, NBR, NDMI, NDSI, and NDWI) and 18 gridded environmental covariates. For training data, they used multi-year field data from BLM Assessment, Inventory, and Monitoring (AIM) data on the North Slope; for assessment data, they compiled field-based measurements of foliar cover from 14 vegetation survey data sets collected between 1998 and 2017.
Of the six species examined, five had models that produced satisfactory maps of foliar cover. Across all models, metrics of elevation, aspect, and surface texture were the most consistently important predictor variables, although other covariates were of differential importance to different species. Interestingly, hydrology-related environmental covariates were generally of low importance, but spectral covariates related to water– specifically NDWI and NDMI– were highly important for most models. In general, composite models for all species overpredicted species absence (foliar cover = 0) at low but significant rates, and underpredicted high values of foliar cover. Specifically, this approach worked well for species that were dominant only infrequently or in specific biophysical contexts, but when applied to species that were usually dominant when present, tended to underestimate cover. Importantly, by comparing their maps to those of a regional categorical vegetation map, Nawrocki et al. demonstrated that their species-level continuous foliar cover maps explained “more of the observed variation in species distribution and abundance (i.e., better represent vegetation patterns).”
Topical Relevance and Methodological Insights:
I found this study by Nawrocki et al. to be an impressively bold but thorough effort to extend the fractional cover approach to individual plant species. To me, their relative success, despite using only moderate resolution imagery, speaks to the power of a number of methodological best practices: namely, the use of (1) multiple spectral indices with demonstrated links to plant physiology, (2) high-quality and spatially proximate field data, and (3) a modelling approach (random forest) that allows for the inclusion of numerous ecologically relevant variables while remaining robust to collinearity. This study also illustrates how the continuous field approach facilitates robust quantification of model performance and estimation of model error, as noted by Coops and Wulder (2019). Their work also reiterates the efficacy of using multi-season imagery (rather than just peak growing season) to discern PFTs in a high-latitude tundra context.
15. Alonzo, M., Dial, R. J., Schulz, B. K., Andersen, H. E., Lewis-Clark, E., Cook, B. D., & Morton, D. C. (2020). Mapping tall shrub biomass in Alaska at landscape scale using structure-from-motion photogrammetry and LiDAR. Remote Sensing of Environment, 245, 111841. [https://doi.org/10.1016/j.rse.2020.111841]
Summary:
In this experiment, Alonzo et al. sought to compare biomass mapping methods for tall shrub vegetation types using two imagery-based structure from motion (SfM) datasets versus airborne LiDAR data, as well as detailed field measurements. In generating SfM data, they trialed both UAV imagery with ground cell resolution of ~ 1.7cm and ~3.1cm hyperspectral imagery from NASA’s G-LiHT Imager, which concurrently collects LiDAR data. Four final model types were assessed: one G-LiHT LiDar-based structure-only model, one G-LiHT SfM-based structure-only model, one G-LiHT SfM-based structure and imagery model, and one UAV SfM-based structure and imagery model. Across all model types, performance was consistently higher for models separated by vegetation type. Important predictor variables were thematically consistent across model types, in that variables related to vegetation height and fractional cover were consistently most important, but models differed in which metrics were most important. Finally, the difference in point density between UAV-derived SfM and G-LiHT-derived (5000–8000 pt/m versus 500–2000 pt/m, respectively), did not significantly impact model fit, meaning that G-LiHT resolution is adequate for shrub mapping where coverage is available.
Topical Relevance and Methodological Insights:
This work by Alonzo et al. was my introduction to structure from motion (SfM) technology, and for me, it critically emphasizes both (1) the validity of using SfM from high resolution imagery in shrub-dominated ecosystems; and (2) the influence of vegetation type on model performance and selection. These experiments were also of interest to me in that they sought to fill a methodological gap in boreal ecology that I’ve observed in the field for decades— namely, that in many inventory and monitoring protocols, tall shrubs are often quantified like an understory layer, even when they are the dominant vegetation. One notable strength of SfM is that, because SfM point clouds are derived from imagery, they have inherent red/green/blue values that improve modeling and classification performance. Interestingly, the importance of vegetation height and fractional cover to final models is consistent with Berner et al. (2018) who were working exclusively with imagery to map biomass.
16. Macander, M. J., Nelson, P. R., Nawrocki, T. W., Frost, G. V., Orndahl, K. M., Palm, E. C., … & Goetz, S. J. (2022). Time-series maps reveal widespread change in plant functional type cover across Arctic and boreal Alaska and Yukon. Environmental Research Letters, 17(5), 054042. [https://doi.org/10.1088/1748-9326/ac6965]
Summary:
The present study represents the kind of change detection analysis emphasized by Coops and Wulder (2019) in their call for continuous field mapping. Building on their fractional cover maps for the North Slope (Macander et al. 2017), Macander et al. (2022) here used 35 years of Landsat imagery and a suite of environmental covariates to map a time series of continuous cover for seven PFTs across Alaska and part of Canada, and then used this multi-temporal map to identify net changes in each PFT over that period. Over the 1,770,000 km2 study area, they identified net increases in all shrub and tree PFTs, net decreases in graminoid and lichen PFTs, and relative stability in the net area of the herbaceous PFT.
One particularly useful consequence of this continuous mapping approach, as highlighted by Coops and Wulder (2019), is that Macander et al. were also able to characterize the spatial distribution of model uncertainty– specifically, by analyzing how root mean square error (RMSE) varied with predicted cover. This analysis indicated that, in general, models underpredicted high cover values for all PFTs, and that model error was minimized for 5 of 7 PFTs at cover values near zero.
Topical Relevance and Methodological Insights:
Because their study area spanned nearly 1.8 million km2, Macander et al. relied on plot data from a number of different field efforts with differing protocols for estimating cover. They specifically chose to model a metric of cover that is directly observable via remote sensing– top cover – and used an empirical model to normalize other estimates of cover to top cover. Different methods of cover estimation are a continual source of uncertainty in longer-term vegetation studies sensu lato, and I found it particularly useful to learn that there are established ways of standardizing plot data across protocols. This approach also addresses, at least partially, the concerns about differences between field-based cover estimation techniques explored by Karl et al. (2017).
This study was also the first of those I read to address spatial autocorrelation. Before generating PFT models, the authors analyzed patterns of spatial autocorrelation between different PFTs and employed a spatially-blocked cross validation approach designed to address such issues. Their model residuals had very little spatial structure, suggesting that this cross-validation approach was successful. These efforts by Macander et al. were an interesting reminder that geospatial and quantitative science are evolving alongside their applied fields, and that new computing and modeling advances continually allow us to tackle previous limitations.
17. Orndahl, K. M., Ehlers, L. P., Herriges, J. D., Pernick, R. E., Hebblewhite, M., & Goetz, S. J. (2022). Mapping tundra ecosystem plant functional type cover, height, and aboveground biomass in Alaska and northwest Canada using unmanned aerial vehicles. Arctic science, 8(4), 1165-1180. [https://doi.org/10.1139/as-2021-0044]
Summary:
Like Nelson et al. (2021), Orndahl et al. (2022) here experimented with the use of high-resolution UAV imagery in mapping high-latitude PFTs, but they also incorporated UAV-derived structural and volumetric data to estimate plant biomass. Specifically, Orndahl et al. coupled UAV-based imagery with UAV-based structure from motion (SfM) scans to generate three-dimensional point clouds that facilitate measurement of plant volumes, from which plant biomass can be estimated. In addition to assessing the accuracy of this approach to biomass modeling, the authors also sought to determine whether vegetation community type influences the accuracy with which a given PFT is modeled.
With regard to imagery-based modeling results, Orndahl et al. had similar outcomes to many others reviewed here: namely that NDVI was the most important classification predictor, and that deciduous and evergreen shrub models had the lowest error (relative RMSE) when evaluated against observed cover values. One particularly interesting result is that SfM canopy height models were much better at discerning canopy-versus-ground points when point clouds were stratified by PFT class, which in turn facilitated improved model performance in complex terrain. As with PFT classification, the authors found that for both canopy height and aboveground biomass, SfM-based models for deciduous shrubs had the lowest RMSE of all PFTs. Interestingly, the authors did not find significant differences in cover prediction error between community vegetation types.
Topical Relevance and Methodological Insights:
Along with Alonzo et al. (2020), this study was my secondary introduction to structure from motion (SfM) technology, and I found its potential to inexpensively address limitations around vegetation structure encouraging. There are some tradeoffs— namely, that SfM point clouds are much denser than those generated by LiDAR and have spectral information, but that SfM cannot penetrate top canopy layers to map underlying vegetation. However, in ecosystems with relatively sparse vegetation, like arctic sedge tundra, SfM is well-suited and cost-effective. It also appears (per Alonzo et al. 2020) to perform reasonably well in more complex tall shrub ecosystems, too. I would be interested to read a comparison of SfM versus LiDAR performance across multiple high-latitude vegetation types to assess complexity-versus-cost and relative performance of each.
18. Nelson, P., Paradis, D., & Hantson, W. (2022). Scaling ground-based hyperspectral scans to AVIRIS next gen using UAV-based VNIR imaging spectroscopy for mapping arctic and boreal plants in Alaska. Authorea Preprints. [https://doi.org/10.1002/essoar.10508608.1]
Summary:
This poster by Nelson et al. reports on work directly addressing issues specific to PFT mapping of arctic tundra vegetation– namely that arctic plant species are often very low-growing and/or highly intermixed. They collected very high spatial- and spectral-resolution data via unmanned aerial vehicle (UAV), and attempted to scale these scans to AVIRIS imagery. The UAV spectrometer they employed collected 326 bands in the visible and near-infrared (VNIR) range and produced ground sample distances (GSD) of 4cm or 10cm. UAV imagery was captured to be coincident with both AVIRIS flights and in-field measurements of PFT reference spectra taken via handheld spectroradiometer.
Nelson et al. successfully compiled 600+ scans from over seventy species into a spectral library and calculated median reflectance by PFT. Though this project included only qualitative accuracy assessments based on optical imagery, Mixture Tuned Matched Filtering (MTMF) appears to have performed better for AVIRIS data than for UAV VNIR data, which the authors attribute to the latter’s narrow spectral range and difficulty with shadows. As this source is a conference poster for then-in-progress work, no finalized results are presented.
Topical Relevance and Methodological Insights:
Nelson et al. (2021) is the first among the sources I reviewed to use a MTMF approach to PFT mapping. MTMF is another method of spectral unmixing that combines elements of the linear spectral mixing model (which, as previously discussed, makes problematic assumptions about end-member constancy) and the Matched Filtering model, which is frequently used in radio and radar applications, but does not translate directly to imaging spectrometry because of fundamental differences between the nature of radio versus optical signals. Both Matched Filtering and MTMF allow for the mapping of a single known end-member without knowing the other background end-member signatures. As such, MTMF provides an alternative to multiple end-member spectral mixture analysis (MESMA) while still accounting for end-member spatial variance.
19. Nelson, P. R., Bundy, K., Smith, K., Macander, M., & Chan, C. (2024). Predicting plants in the wild: Mapping arctic and boreal plants with UAS-based visible and near infrared reflectance spectra. International Journal of Applied Earth Observation and Geoinformation, 133, 104156. [https://doi.org/10.1016/j.jag.2024.104156]
Summary:
This paper reports on the authors’ expanded and finalized work (Nelson et al., 2021, reviewed above) examining the efficacy of unmanned aerial vehicle-based hyperspectral imagery in differentiation of high-latitude PFTs. The authors compared the performance of several machine learning and deep learning approaches to mapping these PFTs, and also explored the importance of spectral resolution in model performance. In general, Nelson et al. found that traditional algorithmic approaches (e.g., random forest) outperformed deep learning approaches. Within the higher-performing models, the most important variables tended to be vegetation-related spectral indices, and models that included wider bandwidth features slightly outperformed those with narrow bandwidth features.
Topical Relevance and Methodological Insights:
This study was the first of the papers I reviewed to mention sampling balance in PFT modeling, which is curious given the frequently heterogeneous nature of arctic and boreal vegetation. The authors took active measures to prevent model bias from PFTs that were proportionally dominant across the study area. Specifically, Nelson et al. built balanced training and testing data sets by randomly sampling pixels until the predetermined sample size had been reached. This seemed, to me, a simple, elegant, and easily replicated solution to a potentially significant source of error.
20. Travers-Smith, H., Coops, N. C., Mulverhill, C., Wulder, M. A., Ignace, D., & Lantz, T. C. (2024). Mapping vegetation height and identifying the northern forest limit across Canada using ICESat-2, Landsat time series and topographic data. Remote Sensing of Environment, 305, 114097. [https://doi.org/10.1016/j.rse.2024.114097]
Summary:
In this study, Travers-Smith et al. (2024) combined multi-temporal composite Landsat imagery with ICESat-2 LiDAR and ASTER DEM data to model tree canopy presence and height, and thereby determine the northern limit of the boreal forest in Canada. Their work provides an elegant example of using multiple remote sensing datasets to examine ecological phenomena from multiple perspectives. Most remote sensing work on treeline advance has focused on vegetation indices derived from multispectral optical sensors, but as LiDAR data improve in spatial and temporal coverage, it has become increasingly feasible to combine imagery and radar data to model horizontal and vertical vegetation structure simultaneously. They used a two-step, random forest classification and modeling approach to first predict canopy presence/absence, and then model canopy height within areas classified as present.
Overall classification (presence/absence) model performance was good, with out-of-bag error values of 8.7% for canopy absence and 5.8% for canopy presence. For this model, the most important variables were median SWIR1 surface reflectance, median tasseled cap brightness (TCB) and median tasseled cap wetness (TCW). The canopy height regression model performed best when canopy heights were between 5-10m, and tended to under-predict canopy values >10m and over-predict canopy values <2m. The most important predictor variables for the regression model included median TCB, median NIR reflectance, and median SWIR2 reflectance. Canopy height predictions generated by these models were consistent with cover values derived by other authors from MODIS and Landsat data. Similarly, Travers-Smith et al. defined “treeline” as the limit of continuous 3m canopy height, which, when mapped, corresponded well with previous models of the forest-tundra boundary derived from aerial photography and plot data.
Topical Relevance and Methodological Insights:
Three of the spectral indices used by Travers-Smith et al.— tasseled cap brightness (TCB), tasseled cap wetness (TCW), and tasseled cap greenness (TCG)— were new to me. The tasseled cap transformation subjects imagery data to principal components analysis, resulting in orthogonal axes, three of which are sometimes split out into indices: brightness, which is a weighted sum of reflectances for all spectral bands and indicates bare soil or rock; greenness, which is associated with green vegetation; and wetness, which is associated with water and soil moisture. The remaining components consist of atmospheric distortions and other image noise related to unimportant optical information. (Secondary source: https://pro.arcgis.com/en/pro-app/latest/help/analysis/raster-functions/tasseled-cap-function.htm).
21. Orndahl, K. M., Berner, L. T., Macander, M. J., Arndal, M. F., Alexander, H. D., Humphreys, E. R., … & Goetz, S. J. (2025). Next generation Arctic vegetation maps: Aboveground plant biomass and woody dominance mapped at 30 m resolution across the tundra biome. Remote Sensing of Environment, 323. [https://doi.org/10.1016/j.rse.2025.114717]
Summary:
Orndahl et al. (2025) used composite Landsat imagery, spatial covariates, and a standardized pan-arctic database of field data to generate machine learning models of plant biomass, woody plant biomass, and woody plant dominance at 30m spatial resolution, along with spatial uncertainty estimates for each. Once these models had been generated, they also examined the effects of bioclimactic zones, growing season temperatures, and vegetation community type on patterns in plant biomass. During model development, the authors compared numerous classification and regression approaches (e.g., linear regression, SVMs, etc) and ultimately decided upon random forest models.
Final maps of biomass density were generated via a presence/absence classification step, followed by regression modeling. For both overall plant biomass and woody biomass models, classification step accuracies were ≥95% and R2 values were ≥ 0.62. Predictor variables with the biggest impact on performance included those derived from the Normalized Difference Water Index (NDWI) and 2-band Enhanced Vegetation Index (EVI2b), as well as bioclimate zone and tree presence. Final models predicted an average woody plant dominance of 41% for the entire Arctic, and mean overall biomass density of ~508gm-2, with significant differences in the mean biomass density between bioclimactic zones. For the Oro Arctic, mean overall biomass density was ~818gm-2 and mean woody biomass density was ~659gm-2, whereas for the High Arctic, predicted means were ~115gm-2 and ~45gm-2, respectively.
Biomass density was similar between general vegetation community types (wetland v. graminoid v. shrub tundra), but exhibited stronger patterning at finer levels of classification; for example, low shrub tundra exhibited significantly higher biomass than any dwarf shrub classes. Other spatial patterns in biomass distribution included regional hotspots of biomass in western Alaska, the Mackenzie River delta, and the Nenets region of Russia, as well as local hotspots along riparian corridors and in some–but not all– old burn scars. Finally, Orndahl et al. found that both mean overall and mean woody plant biomass consistently increased with thawing degree days (their chosen growing season temperature metric), as did woody plant dominance.
Topical Relevance and Methodological Insights:
To me, this study represents a benchmark in the new generation of remote sensing-derived continuous field vegetation modeling, as first envisioned by DeFries et al. (1995) three decades ago. When DeFries et al. (1999) first mapped continuous field vegetation parameters with AVHRR, MODIS was still a future mission! But in the introduction to the present paper, Orndahl et al. contextualize their work by describing the shortfalls of previous biomass modeling efforts derived from coarse (AHVRR) and moderate (MODIS) resolution sensors. They highlight the heterogeneity of arctic communities and the necessity of both sufficient field data and higher resolution imagery to adequately capture this spatial variability. As a part of their broader work on this topic, they also developed a custom script to convert R-based model fits to Google Earth Engine code, which I will definitely be bookmarking for future use!
22. Steckler, M.R., Kumar, J., Breen, A.L., Zhang, T., Hoffman, F.M., Hargrove, W.W., Walker, D.A., Wells, A.F., Droghini, A., Nawrocki, T.W. and Wullschleger, S.D., 2025. PaVC: the foundation for a Pan-arctic Vegetation Cover database. Scientific Data, 12(1), p.1271. [https://doi.org/10.1038/s41597-025-05326-9]
Summary:
This essential work by Steckler et al. details recent efforts to compile a central, standardized database that houses field data from various projects across the arctic and can be used in vegetation mapping applications. At time of publication the nascent (PaVC) database contained standardized data for only Alaska, but efforts to expand coverage to different countries and languages are underway. While simple in concept, this effort entailed substantial effort to synonymize disparate data formats, methodologies, and taxonomic standards. To achieve this goal, the authors developed an open-source Python workflow capable of synthesizing cover data collected via visual estimation or point-intercept methods and of summarizing cover by individual species or by PFT. The tables that define accepted taxonomies and PFT classification can also be edited by the user to fit their schema and project needs. They also included ancillary data fields on collection date, methodology, project purpose, author(s), and other attributes by which it might be useful for end users to filter records. The authors specified that additional filterable fields are being added for future iterations of the database, too.
Topical Relevance and Methodological Insights:
This effort by Steckler et al. provides an extremely valuable resource for vegetation model training and validation that can easily be filtered and/or scaled by the end user according to their data needs. Many of the papers I’ve reviewed emphasize the importance of high-quality, spatially proximal field data for predictive model performance. The value of the nascent PaVC lies not only in its immediate applicability to remote sensing analyses but also in its open-source design and reproducibility.
23. Zhang, T., Steckler, M. R., Breen, A. L., Hoffman, F. M., Hargrove, W. W., Salmon, V. G., … & Kumar, J. (2025). Mapping wall-to-wall fractional cover of Arctic tundra plant functional types in Alaska using 20-m spatial resolution satellite imagery and harmonized plot observations. International Journal of Applied Earth Observation and Geoinformation, 144, 104892. [https://doi.org/10.1016/j.jag.2025.104892]
Summary:
This very recent paper maps six PFTs across the Alaskan Arctic at the highest spatial resolution yet, using a unique combination of data types and with remarkable accuracy for all PFTs. Specifically, Zhang et al. used a random forest modeling approach to map fractional cover at 20m resolution from three data sets: multispectral imagery from Sentinel-2, synthetic aperture radar (SAR) data from Sentinel-1, and a digital elevation model (DEM) from the ArcticDEM mosaic. From these data sets, the authors compiled more than 30 variables for model input, including two SAR bands, ten reflectance bands and 19 spectral indices, and DEM-derived values of elevation, slope, and aspect.
The models for non-vascular and bryophyte PFTs performed remarkably well, and were only outdone by those for evergreen and deciduous shrubs. In general, final maps corresponded well with both fractional cover estimates generated by Macander et al (2017) using Landsat imagery, and with dominance classes depicted by the Circumpolar Arctic Vegetation Map (CAVM; Walker et al. 2005). Interestingly, the authors noted improved model discrimination between graminoids and shrubs in the present analysis, relative to that of Macander et al. (2017), which they attribute both to plot-QC methodology and to the absence of scan line artifacts introduced via Landsat imagery. Importantly, though, the authors undertook a “leave-one-site-out” cross validation approach that emphasized the importance of field data and vegetation structure in model performance: for example, the final overall model for graminoids had a correlation coefficient of R2 =0.75 and RMSE = 0.08, but when sites from a project that heavily sampled graminoid tundra were excluded, model performance dropped to R2 = 0.04. Finally, the models developed here consistently struggled with the litter PFT (best R2 = 0.36 and RMSE = 0.07), which the authors attributed to the temporal variability.
Topical Relevance and Methodological Insights:
This paper was conceptually interesting to me in its relatively anomalous use of Sentinel data rather than Landsat data, the latter of which has strongly dominated among the recent studies I’ve reviewed. Sentinel-2 imagery contains a quality assessment band (QA60) analogous to Landsat’s QA_PIXEL band for use in cloud masking, in addition to providing significantly higher resolution. Zhang et al. also provided two valuable methodological insights with regard to data collection inconsistencies within the Pan-Arctic Vegetation Cover database: First, they addressed plot size differences by averaging cover values of larger plots and aggregating smaller ones into “parent plots” of similar size. Second, they employed an outlier detection approach to account for anomalous and low-quality plot data not otherwise accounted for during harmonization efforts by Steckler et al. (2025).
Other Resources
USDA-ARS, 2025. Remote Sensing Methods. The Landscape Toolbox. Accessed online 25 October 2025. [https://landscapetoolbox.org/remote-sensing-methods/]
Summary:
The Landscape Toolbox is a fabulous resource I first learned about through my involvement in the Bureau of Land Management’s AIM program. The Landscape Toolbox is a multi-partner project between BLM AIM, the USDA-ARS Jornada Experimental Range, and Idaho’s Nature Conservancy chapter, and it provides a publicly accessible collection of ecological analysis and monitoring tools, protocols, and resources relevant to scientifically rigorous landscape management. Throughout this project, I found myself referring back to it periodically to keep various technical definitions and acronyms straight, and so have included it here and highlighted a few key definitions.
Definitions of Interest:
– Spectral Mixture Analysis (SMA, aka subpixel analysis) is a technique for estimating the proportion of a mixed pixel that is covered by a series of known cover types. SMA predicts the proportion of a pixel that belongs to a particular class based on known spectral characteristics of its end-members. For vegetation mapping, ‘end-member’ refers to the spectra that correspond to “pure” (unmixed) pixels of a given land cover class, which makes both the definition of land cover classes, and selection of appropriate end-members for each of these classes, critical to successful SMA. One solution is to obtain end-members from the analysis image itself, so that no further calibration is needed.
– Multiple End-member Spectral Mixture Analysis (MESMA) is a version of SMA that addresses the problem of spatial dependence in end-member selection, by allowing the number and the quantity of end-members to vary pixel by pixel. Traditional SMA models pixel spectra as a linear combination of spectral signatures from a set of surface features which are assumed not to vary across the image. SMA is not equipped to deal with scenarios in which a given class is absent from a pixel, nor the reality that end-member reflectance can vary within and between images/scenes. MESMA allows for variation within the set of considered end-members, and estimates proportional cover by using whichever model has the smallest RMSE when compared to the spectral curve of the pixel.