Introduction
Developing a plot-based, field sampling design to quantify surface properties of remotely sensed pixels involves a trade-off between implementation efficiency and sample error. Intuitively, collecting either more or larger plots is expected to reduce sampling error. However, optimizing the balance between efficiency and error requires quantification of how error varies according to the number, size, and configuration of plot samples used to capture the spatial variability of the surface property of interest for a region. For retrieval of surface properties, such as lichen cover, from moderate spatial resolution remote sensing sensors (e.g., 10–30 m), it is also important to understand how accuracy is influenced by field plot sample error and the sample size of pixels used for model training. Here, we examine sampling effects as they pertain to field measurements and the retrieval of surface cover properties from remotely sensed images. To avoid ambiguity in the terminology we use related to sampling, a list of definitions is provided in Table 1.
[Image omitted: See PDF]
The effectiveness of different sampling methods has been widely studied for many vegetation measurements, among other ecosystem characteristics (Goslee 2006; Symstad et al. 2008; Bonham 2013; Pilliod and Arkle 2013). However, determining the optimal sampling for retrieval of a specific surface property, at a given scale, and across a highly variable region is not easily deduced to guide sampling design. Images collected using a remotely piloted aircraft system, or drone, provide a new means to address this issue, and have been used to investigate scaling related to the remote sensing of vegetation properties at northern latitudes (Assmann et al. 2020; Siewert and Olofsson 2020). In particular, drone-based mapping appears well suited to studying field plot and satellite pixel sampling error, since its very high resolution and spatial continuity can be used to (a) simulate a wide range of field plot sampling designs and (b) quantify the actual surface cover within individual satellite-sensor pixels.
There are a few recent studies that have used drone data to investigate error of different plot-based, field sampling strategies. One study used single drone images to derive the fraction of green vegetation cover contained within a 10 × 10 m area designed to be representative of a Sentinel-2 pixel footprint (Chen et al. 2019). A varying number of 0.5 m quadrat plots were then individually captured by ground-level photographs and classified within this 100 m2 area to determine the ability to estimate the drone-derived vegetation cover. Another investigation (Hao et al. 2021) captured nested drone surveys covering 100 ha at 2.7 cm spatial resolution to classify desert vegetation/non-vegetation, and a second covering 1 ha at 0.4 cm spatial resolution to classify two desert species of interest. They then examined the effect that different plot sizes (6–500 m) covering the survey area had for predicting plant species density. Finally, in a similar study by Hao et al. (2020), optimal quadrat plot sizes for sampling the cover of sparse desert vegetation within seven 100 × 100 m plots were determined using drone-based mapping. By analyzing the frequency distribution and variation of vegetation cover derived from Monte Carlo sampling of single 10–60 m plots, they found that a 50 m extent provided the most appropriate sampling for the study location.
In northern regions, evidence suggests that climate change has reduced lichen cover or volume due to increased fire occurrence (Joly et al. 2009; Hu et al. 2015; Tsuyuzaki et al. 2018) and competition from vascular plants (Tremblay et al. 2012; Fraser et al. 2014; Moffat et al. 2016). This has motivated research for monitoring lichen properties because many species of lichen are an important food source for caribou, particularly during the winter months (Llano 1956;Jandt et al. 2008; Joly et al. 2010, 2015; Joly and Cameron 2018). Satellite image-based modeling of reindeer lichen (Cladonia spp.) cover or volume, underpinned by field plot reference data, is becoming a common method for quantifying caribou lichen forage over large spatial extents (Theau and Duguay 2004a, b; Nelson et al. 2013; Falldorf et al. 2014; Kennedy et al. 2020; Macander et al. 2020). High-resolution drone mapping has recently been used to provide more comprehensive reference data for training satellite-based lichen retrieval models, assuming that it provides more accurate sampling of satellite pixels compared to conventional field plots (Macander et al. 2020; Fraser et al. 2021; He et al. 2021). No study has explicitly examined the effectiveness of different plot-based sampling approaches for providing representative reference data under different satellite resolutions and types of lichen habitats. Recently collected drone surveys from woodlands in Newfoundland and Labrador and from rocky Canadian Shield in the Northwest Territories (NWT), Canada, provided an opportunity to examine this issue by using them to simulate plot measurements from 2 to 3 cm resolution lichen classifications.
Objectives
The purpose of this study was to use drone-based lichen maps to quantify plot sampling error and its effect on lichen cover retrieval with moderate spatial resolution remote sensing sensors. Lichen cover was used as a case study in this analysis, but the methodology could be applied to other ground cover types. Specifically, the objectives of this research were to:
1. Quantify the error resulting from different simulated quadrat field plot sampling scenarios when they were used to estimate lichen cover within a collection of remotely sensed pixel footprints. These scenarios examine a range of plot sizes, number of plots, and sampling strategies.
2. Evaluate the effects of plot sampling error (objective 1) and the number of pixels used for training on lichen cover retrieval accuracy with moderate spatial resolution imagery.
Data and methods
Drone-based lichen maps
For the plot-simulation scenarios, existing drone-based lichen surveys acquired during the summer of 2019 were used from two unique, lichen habitats located along the Ingraham Trail near Yellowknife, NWT, and in the Labrador region of Newfoundland and Labrador, Canada (Fig. 1). Two sites surveyed in NWT were in rocky uplands with low conifer tree density, where lichen cover occurred in discontinuous patches generally less than 0.25 m2. The drone maps were developed to represent forage lichens largely composed of fruticose growth forms and Cladonia species. Crustose and some foliose lichens growing on rocks were also abundant at some sites, but were not included in these lichen cover maps. Drone images were acquired using a DJI Matrice 200 quadcopter carrying a Micasense Altum multispectral camera that captured five visible to near-infrared bands at 3.3 cm spatial resolution. The lichen maps were derived using Pix4d structure-from-motion (SfM) processing and random forest (RF) classification of the resulting image mosaics and had an accuracy greater than 95% (Fraser et al. 2021). Two drone survey sites in Labrador consisted of lichen woodlands, where thicker and more continuous lichen cover was broken up by tree cover, downed woody debris, and water. SfM orthomosaics were generated using a DJI Mavic 2 Pro, equipped with the standard RGB camera and an additional Sentera camera, capturing visible bands at 2 cm spatial resolution and red-edge and near-infrared bands at 6 cm spatial resolution. Only the near-infrared band was used from the Sentera camera resampled to 2 cm. Lichen maps were then derived using support vector machine classification that yielded an accuracy greater than 93% (He et al. 2021).
[Image omitted: See PDF]
Moderate spatial resolution satellite image data
To investigate the effects of field plot sample error on machine learning retrieval models of lichen cover at moderate spatial resolution, Sentinel-2 and Landsat-8 images were acquired as summarized in Table 2. For Sentinel-2, all 10 m bands (blue, green, red, near-infrared) were used plus the shortwave bands at 20 m spatial resolution. For Landsat-8, all bands except for the coastal blue, panchromatic, and thermal bands were used.
[Image omitted: See PDF]
To ensure accurate geolocation between the drone orthomosaics and moderate spatial resolution data, an exhaustive image matching strategy was used. For this, shifts of 1 m were applied to the vertical and horizontal satellite image coordinates and compared to the drone imagery resampled (averaged) to match the shifted grid. The correlation between these was recorded for shifts covering a full pixel. Once the optimal shift was determined for the 1 m interval, it was refined using a 10 cm interval. Further refinements did not improve results. Shifts in the range of 0.2–8 m were determined for Sentinel-2 and 0.1–12 m for Landsat-8. Comparison of lichen cover retrieval before and after applying shifts showed significant improvements in accuracy.
Assessment of field plot sampling error
To assess the error in lichen cover for a sensor pixel associated with different field plot sampling scenarios, we examined spatial resolutions of 10 and 30 m representing Sentinel-2 and Landsat-8 sensors, respectively. In addition, 6 m spatial resolution was also included to assess sampling effects for a smaller spatial unit following Fraser et al. (2021).
The plot sampling scenarios varied the plot size, number of plots, and sampling strategy used to estimate the cover for a sensor pixel. An overview of the analysis undertaken is provided in Fig. 2. Plot sizes simulated from the drone lichen classifications consisted of 0.5, 1, and 2 m using a square or quadrat sampling unit. The number of sample plots ranged from one to the maximum number that could be contained within the sensor pixel without overlap. For the sampling strategies, both random and systematic approaches were tested. Figure 2 shows the grid for a sensor with 30 m spatial resolution overlaid on the extent of the drone-based lichen classification. Within a sensor-grid cell (i.e., sensor pixel), a nested plot-grid was defined for each of the plot sizes investigated. Cells from the plot-grid were randomly sampled and used to estimate the sensor-grid cell lichen cover. For systematic sampling, equally spaced plot sets were defined such as 2 × 2, 3 × 3, 4 × 4, etc. within a sensor pixel. From the sampled plots, the cover for a pixel was estimated as the sum of cover in the plots multiplied by the ratio of the sensor pixel to the total plot area.
[Image omitted: See PDF]
The estimated cover from the plots was compared to the true lichen cover from the drone image classification within a sensor pixel. The absolute error was computed between these for all pixels falling within the drone survey extents. As results were similar between sites in the NWT or Labrador regions, all errors were pooled to compute the 95th percentile error. This was done to summarize and present the results more effectively.
Effect of field plot sample error and training sample size on lichen cover retrieval
The influence of error related to plot sampling (plot size and number of plots) and sample size (number of sensor pixels sampled for training) on modeling lichen cover at moderate spatial resolution was investigated using the Sentinel-2 and Landsat-8 images listed in Table 2. For this assessment, training data were collected from one drone survey area and the western half of the second survey area within each region. The remaining eastern half of the survey area was used as a spatially independent test set. Training data collection from the two surveyed areas and each region provided for a total of four comparisons. This holdout approach was used to develop a robust model and to examine the potential for short-distance extrapolation, which is more likely how drone data would be used for moderate resolution mapping applications over large spatial extents. For testing, the same pixel set and cover values from the drone classification were used and did not change between sampling scenarios and model comparisons.
RF and Gaussian process regression (GPR) machine learning models were both assessed for lichen cover modeling using the scikit-learn module in Python (Pedregosa et al. 2011). Inputs consisted of the spectral bands from Sentinel-2 or Landsat-8, and cover was the modeled response in the range 0 to 1. For both models, numerous parameters were investigated. The final set for RF was the number of trees set to 1000, the number of features considered at each split was set to the square root of the total features, and the minimum samples per leaf was set at 2. For GPR, several kernels were tested including Radial-basis, Matérn, and rational quadratic with or without an additive white noise kernel. Kernel parameters were optimized in scikit-learn by maximizing the log-marginal-likelihood with 10 restarts.
To evaluate results, the modeling efficiency coefficient (MEC), mean error (ME), mean absolute error (MAE), root mean square error (RMSE), and the coefficient of determination R2 were computed. The MEC and R2 are similar measures, but the MEC can be negative indicating the mean provides a better model. The R2 in this analysis includes a linear recalibration of the model prediction to the observed results and was added to allow for comparison with other studies that report this metric.
To evaluate the relative effect of plot sampling error on lichen cover retrieval models, systematic plot configurations from a single plot (1 × 1) in the center of the sensor pixel to 8 × 8 plots were evaluated. For the first analysis, all sensor pixels were used. In the second, for a subset of systematic plots (1 × 1, 2 × 2, 3 × 3, 4 × 4), sensor pixels were sampled at selected intensities (sample size) between 1% and 80% of the available training pixels for each sampling comparison.
To compare the effects of the different sampling factors on the lichen cover retrieval accuracy, a multiple linear regression model was developed that included plot size, number of plots, sample size in %, sensor, and site as independent variables. Numeric variables were standardized so that model coefficients indicate the relative effects of each sampling factor. The sensor and site variables were coded as indicator variables with two levels: Sentinel-2 and Landsat-8 for sensor, and NWT and Labrador for site. Two models were developed for this purpose, one using all results and a second where the training sample size was less than 15 pixels and the 1 × 1 single central plot results were removed. The second model was included to evaluate the effect on accuracy over a more realistic set of sampling conditions.
Results
Assessment of field plot sampling error
The results of the sample plot analysis for the different sensor pixel sizes (6, 10, and 30 m), regions, and sampling configurations are shown in Fig. 3. In all cases, the error followed a negative power function, dropping rapidly with the number of plots to the range between 5 and 15 plots, and more slowly afterwards. A comparison of random and systematic sampling showed that systematic provided reduced error in most cases. However, when the sampled plot area was low relative to the pixel area (sum of plot area/pixel area), there was little difference, both producing high error because neither sample was representative. When the sampled plot area was high, the differences were also small because the plots covered a large proportion of the pixel and thus both provided a good estimate. The advantage of systematic sampling therefore occurred between the low and high sampled area (which is a function of the number of plots and plot size) because it was more likely to provide a representative estimate when the cover within the sensor pixel was spatially variable. Random sampling has a higher probability of casting plots that could capture too much or too little cover when the cover was highly spatially variable or clustered.
[Image omitted: See PDF]
To allow for easier comparison between regions, plot sizes, and sensor resolutions, the number of plots needed in each case to achieve an allowable error of 10% for 95% of the sampled pixels is summarized in Fig. 4. Three main findings are evident in this figure. First, more plots were clearly required in Labrador compared to NWT. Second, increasing the plot size decreased the number of plots needed as expected, but the effect was stronger for the Labrador site. Third, the NWT sites were more invariant to sensor resolution compared to Labrador.
[Image omitted: See PDF]
These results were driven by the spatial variance or patchiness of lichen cover at the plot and sensor pixel scales. In the case of NWT, a more uniform distribution of lichen cover occurred at both scales. By comparison, in Labrador, the lichen cover was more clustered in larger patches controlled dominantly by tree and water cover. These properties of the two study regions can be seen in Fig. 1. To better quantify these results, the spatial variability of the lichen cover for a plot size of 1 m within a 10 m sensor pixel was computed for pixels that overlapped the drone lichen cover map. This involved extracting the cover for all cells in the 1 m grid for a pixel as depicted in Fig. 2 (yellow grid) and computing the standard deviation from these. If the lichen cover distribution within the pixel was spatially uniform or well dispersed then the standard deviation will be low. If it was clustered then it will be high as plots will have a greater mix of cover values. Figure 5 shows a comparison of the cover standard deviation and observed error (based on 1 m plot size and 3 × 3 systematic plots) for pixels in each site. As the standard deviation increases so too did the error bounds. A higher standard deviation may not lead to higher error in all cases due to the statistical nature of sampling. A comparison of the two regions shows that the Labrador sites had greater spatial variance and thus error for this sampling scenario.
[Image omitted: See PDF]
Effect of field plot sample error and training sample size on lichen cover retrieval
The benchmark cover accuracy for each site comparison and machine learning model are shown in Table 3. Here the sample size was 100% of the available sensor pixels and the cover values were extracted from the drone classification, not sampled using plot simulations. Thus, these results provide an upper bound accuracy for each site comparison. Of the two machine learning methods, RF produced higher accuracies compared to GPR in most cases so it was used for further analysis. GPR can produce excellent results, but we found its optimization less consistent compared to RF in some tests leading to reduced accuracy.
[Image omitted: See PDF]
The effect of the number of field plots and plot size on RF accuracy using the full set of satellite pixels for training are shown in Fig. 6. The use of a single plot was not effective for any plot size as expected. For plot sizes of 1 m or larger, the 3 × 3 systematic plot configuration was close to the upper bound accuracy in most cases and larger plot sets of 4 × 4 to 8 × 8 improved on this only marginally. The results for Landsat-8 were more variable and suggest that the 6 × 6 plot configuration was needed to be close to the highest accuracy. For the 0.5 m plot size, using the 3 × 3 plot configuration moderately reduced accuracy for both sensors and 6 × 6 plots were needed to avoid a reduction in accuracy.
[Image omitted: See PDF]
The combined effect of the field sample error for a pixel and the number of pixels sampled for training a cover retrieval model is also an important consideration related to the sampling efficiency/error trade-off. Results for the 1 m plot size, systematic plot configurations, and number of pixels sampled are shown in Fig. 7. The results for Sentinel-2 with 3 × 3 systematic plots or more showed an initial rapid increase in accuracy to around 200 samples, then a slow linear increase. More than 500 samples were needed to be within a few % of the upper bound accuracy. This represented approximately 60% of the sensor pixels for NWT and 20% for the Labrador data sets. The change in plot size reduced variation observed between the systematic plot scenarios but retained the same properties noted above.
[Image omitted: See PDF]
For Landsat-8, the results were more variable due to two related factors. The first was the number of pixels sampled was reduced by a factor of nine relative to Sentinel-2. Second, the aggregation of cover within a 30 m pixel reduced the variability of cover values across pixels. This can lead to low or negative MEC results because the mean can become a sufficient predictor for a small data range, which was the case for NWT. Despite this, the results for Landsat-8 point to similar conclusions to those for Sentinel-2, where accuracy increases quickly to about 40 samples followed by a slow linear increase with more samples. In terms of sample %, this was approximately the same as the Sentinel-2 results of 60% for NWT and 20% for Labrador.
The relative effects on cover retrieval accuracy of sampling factors (field plot size, number of field plots, number of training samples), site, and sensor from the regression analysis are shown in Fig. 8. The first model used all possible results and had an R2 value of 0.38. The second model removed sample sizes less than 15 pixels and the single field plot (1 × 1) results, producing a larger R2 of 0.6. In both models, statistical variability of RF accuracy was high for small training sample sizes, few field sample plots, and small plot sizes resulting in larger residuals, but was much more pronounced for the first model. Both show the same relative results, with the sensor and site indicating larger effects. Of the sampling factors, the number of training samples had the biggest influence and plot size the least.
[Image omitted: See PDF]
Discussion
The results of the drone-based, plot simulation analysis provide quantitative insight into field sampling design in similar environments or where extrapolation of the results would be reasonable. The relative influence of the sampling factors investigated here is also expected to be similar for comparable cover types. Here, we examined two different spatial distributions of lichen cover: low cover with a uniform distribution in NWT and high cover with a more clustered distribution in Labrador. These were representative of lichen cover in these regions for open forest cover conditions. Thus, depending on the expected or estimated cover in other regions based on previous fieldwork or high-resolution image sources, a conservative sample design could be developed that is informed by the results of the analysis presented here. At more northern latitudes, lichen cover tends to be lower and more discontinuous and may not be directly observable at the drone spatial resolutions examined here (Kennedy et al. 2020). This may also be the case in regions where significant foraging and trampling of lichen has occurred. Repeating this type of analysis over a larger range of regions and cover types would help define general guidelines for sampling to support remote sensing of ground cover distributions.
Alternatively, the drone-based methodology can be implemented to determine an optimal sampling design. There are countless sampling designs that can be considered that were not addressed in this study. Drone collection is advantageous as other sampling designs can be implemented, evaluated, and directly compared. Drone maps can also be used to estimate the spatial variance at different scales of field plots or sensor pixels. This can be used with statistical methods to estimate the number of plots required for a pixel in a more adaptive manner (NIST/SEMATECH 2020). Such an approach could be useful where an in-site visit is desired to support more detailed vegetation descriptions. As an example, it is often desirable to measure volume/biomass of lichen in addition to cover. However, these measurements can be costly to acquire and are most accurately obtained in situ. For this reason, statistical relationships are often developed between cover and volume/biomass, which are then used with more extensively measured cover data (reviewed in Greuel et al. 2021).
In this study, the factors (shown in Fig. 8) with the largest impact on lichen cover retrieval accuracy were the spectral separability of lichens relative to other covers (site effects) and the spatial resolution of the sensor (Landsat-8 or Sentinel-2). The number of pixels sampled for training was the most significant of the sampling-related factors. In this case, where drones can capture the cover type of interest, they provide a means to acquire a large amount of training data. However, a potential drawback of drone data collection is spatial autocorrelation. Numerous remote sensing studies have highlighted the potential reduced model performance or biases associated with spatial autocorrelation in training data or between training and test data (Dobbertin and Biging 1996; Chen and Wei 2009; Millard and Richardson 2015; Kennedy et al. 2020; Rocha et al. 2021). A potential advantage of field plots in this regard is they can be more widely distributed over the study area to reduce autocorrelation, but fewer plots will be acquired. Another consideration is that field sampling may be needed as a backup in remote locations if drones cannot be used due to weather or other flight constraints. The specific trade-offs regarding autocorrelation and sample size can also be studied using the approach developed here.
Intuitively, more and (or) larger plots will provide more accurate cover estimates, but depending on the field collection method may be impractical to implement due to cost. For example, point density sampling for a large plot can be very time consuming and results here do not suggest that increasing the plot size is as important as the number of plots. Visual methods are also often used for lichen cover estimation in the field as they are highly efficient to implement (e.g., Moen et al. 2009; Rosso et al. 2014). Visual estimation can be easily adapted to larger plots but could lead to reduced precision. Based on the result of this analysis, a 1 m plot size using a systematic sample of 3 × 3 plots offered a good compromise between error and effort for most scenarios (Figs. 6 and 7). However, for Landsat-8, increasing the number of plots is likely necessary where the cover variability is large or clustered at the 30 m spatial scale. Systematic sampling was generally found to produce lower field plot sample error, but this was dependent on the sample area intensity. For very low sampled area (small plot size and few plots) or very large sampled area (larger plot size and many plots), systematic and random sampling were similar. Systematic sampling is generally easier to implement in many field environments as each plot does not have to be geolocated. Only the first plot needs to be precisely located and the others measured in relation to it. For sites with taller and denser vegetation, systematic sampling may not provide an efficiency advantage due to difficulty of locating plots in these conditions.
Geolocation uncertainty is another factor to consider when developing machine learning satellite remote sensing cover models. Ensuring that plots fall completely within a pixel in the field can be difficult, requiring prior knowledge of the sensor grid and ability to precisely geolocate field sample plots. Here we refined the geolocation between the moderate spatial resolution and drone images. This is an advantage of drone mapping because it allows for adjustment post-collection, whereas for field plots this is more difficult or not possible. Clearly, as plots fall outside the bounds of the sensor footprint error may increase. This effect will be tempered with more spatially uniform cover distributions and slightly by the modulation transfer function of the sensor, which includes some of the area outside the pixel in the recorded signal. Residual atmosphere effects may also be a minor factor.
Landsat-8 lichen cover retrieval showed lower accuracy compared to Sentinel-2 with respect to the MEC and R2. Comparing the MAE between Landsat-8 and Sentinel-2 showed similar results. Over larger extents, Landsat results are likely to improve as the sample size and range of cover values increase. At coarse scales, cover values tend to be less variable due to spatial aggregation and this can improve accuracy relative to finer scales if a sufficient data range is maintained. Dark and Bram (2007) reviewed this in the context of the modifiable area unit problem. Recent scaling studies in northern environments have highlighted this for green vegetation cover (Riihimäk et al. 2019) and lichen cover (Fraser et al. 2021).
As a final note, the results of this study could be used to inform a weighting scheme for incorporating diverse observations in remote sensing cover models. There are numerous data sources that can be leveraged for model development. For example, there may be historical measurements available for a sensor pixel based on a single 1 m point sample frame and visual cover estimates based on several 1 m plots. The results here help to inform the relative weights that should be assigned for single vs. several plots in addition to knowledge of the sample accuracy and precision factors related to the collection method.
Conclusions
For the development of remote sensing-based lichen cover models, the number of pixels sampled (using field plots) for training was observed to have a greater effect on accuracy than plot measurement factors such as plot size and number of plots sampled within pixel footprints. Large differences in accuracy were also found between study sites due to spectral confusion between lichen and other cover types (e.g., rock) and sensor resolution. For most cases, a 1 m plot size using a systematic sample of 3 × 3 plots offered a good compromise between error and effort, but additional plots maybe required for more spatially variable cover and at larger sensor spatial resolutions. The results provide quantitative insight into field sampling that can be used to inform sampling designs for other studies. Alternatively, the method can be replicated to better quantify and optimize a sampling design for retrieving surface properties of a given cover type for a region.
Acknowledgements
We would like to recognize the support from Jason Duffe and Jon Pasher regarding the funding and administration of this research at Environment and Climate Change Canada and the Natural Resources Canada Program, Earth Observation Baseline Data for Cumulative Effects (EO4CE) led by Darren Jansen.
Assmann J.J., Myers-Smith I.H., Kerby J.T., Cunliffe A.M., Daskalova G.N. 2020. Drone data reveal heterogeneity in tundra greenness and phenology not captured by satellites. Environmental Research Letters, 15: 125002:1–15.
Bonham C.D. 2013. Measurements for terrestrial vegetation. 2nd ed. John Wiley & Sons, Hoboken, NJ.
Chen D., Wei H. 2009. The effect of spatial autocorrelation and class proportion on the accuracy measures from different sampling designs. ISPRS Journal of Photogrammetry and Remote Sensing, 64: 140–150.
Chen J., Zhao X., Zhang H., Qin Y., Yi S. 2019. Evaluation of the accuracy of the field quadrat survey of alpine grassland fractional vegetation cover based on the satellite remote sensing pixel scale. ISPRS International Journal of Geo-Information, 8: 497.
Dark S.J., Bram D. 2007. The modifiable areal unit problem (MAUP) in physical geography. Progress in Physical Geography: Earth and Environment, 31(5): 471–479.
Dobbertin M., Biging G.S. 1996. A simulation study of the effect of scene autocorrelation, training sample size and sampling method on classification accuracy. Canadian Journal of Remote Sensing, 22(4): 360–367.
Falldorf T., Strand O., Panzacchi M., Tømmervik H., 2014. Estimating lichen volume and reindeer winter pasture quality from Landsat imagery. Remote Sensing of Environment, 140: 573–579.
Fraser R., Olthof I., Kokelj S., Lantz T., Lacelle D., Brooker A., 2014. Detecting landscape changes in high latitude environments using Landsat trend analysis: 1. Visualization. Remote Sensing, 6: 11533–11557.
Fraser R.H., Pouliot D., van der Sluijs J. 2021. UAV and high resolution satellite mapping of forage lichen (Cladonia spp.) in a rocky Canadian Shield landscape. Canadian Journal of Remote Sensing, 48: 5–18.
Goslee S.C. 2006. Behavior of vegetation sampling methods in the presence of spatial autocorrelation. Plant Ecology, 187: 203–212.
Greuel R.J., Degré-Timmons G.É., Baltzer J.L., Johnstone J.F., McIntire E.J.B., Day N.J., 2021. Predicting patterns of terrestrial lichen biomass recovery following boreal wildfires. Ecosphere, 12(4): e03481.
Hao L., Shi Q., Imin B., Kasim N. 2020. Methodology for optimizing quadrat size in sparse vegetation surveys: a desert case study from the Tarim basin. PLoS ONE, 15(8): e0235469.
Hao M., Zhao W., Qin L., Mao P., Qiu X., Xu L., 2021. A methodology to determine the optimal quadrat size for desert vegetation surveying based on unmanned aerial vehicle (UAV) RGB photography. International Journal of Remote Sensing, 42(1): 84–105.
He L., Chen W., Leblanc S.G., Lovitt J., Arsenault A., Schmelzer I., 2021. Integration of multi-scale remote sensing data in reindeer lichen fractional cover mapping in eastern Canada. Remote Sensing of Environment, 267: 112731.
Hu F.S., Higuera P.E., Duffy P., Chipman M.L., Rocha A.V, Young A.M., 2015. Arctic tundra fires: natural variability and responses to climate change. Frontiers in Ecology and the Environment, 13: 369–377.
Jandt R.R., Joly K., Meyers C.R., Racine C. 2008. Slow recovery of lichen on burned caribou winter range in Alaska tundra: potential influences of climate warming and other disturbances. Arctic, Antarctic and Alpine Research, 40: 89–95.
Joly K., Cameron M.D. 2018. Early fall and late winter diets of migratory caribou in northwest Alaska. Rangifer, 38(1): 27–38.
Joly K., Chapin F.S., Klein D.R. 2010. Winter habitat selection by caribou in relation to lichen abundance, wildfires, grazing, and landscape characteristics in northwest Alaska. Écoscience, 17: 321–333.
Joly K., Jandt R.R., Klein D.R. 2009. Decrease of lichens in arctic ecosystems: the role of wildfire, caribou, reindeer, competition and climate in north-western Alaska. Polar Research, 28: 433–442.
Joly K., Wasser S.K., Booth R. 2015. Non-invasive assessment of the interrelationships of diet, pregnancy rate, group composition, and physiological and nutritional stress of barren ground caribou in late winter. PLoS ONE, 10(6): e0127586.
Kennedy B., Pouliot D., Manseau M., Fraser R., Duffe J., Pasher J., 2020. Assessment of Landsat-based terricolous macrolichen cover retrieval and change analysis over caribou ranges in northern Canada and Alaska. Remote Sensing of Environment, 240: 111694.
Llano G.A. 1956. Utilization of lichens in the Arctic and subarctic. Economic Botany, 10: 367–392.
Macander M.J., Palm E.C., Frost G.V., Herriges J.D., Nelson P.R., Roland C., 2020. Lichen cover mapping for caribou ranges in interior Alaska and Yukon. Environmental Research Letters, 15: 055001.
Millard K., Richardson M. 2015. On the importance of training data sample selection in random forest image classification: a case study in peatland ecosystem mapping. Remote Sensing, 7: 8489–8515.
Moen J., Danell Ö., Roger H. 2009. Non-destructive estimation of lichen biomass. Rangifer, 27(1): 41–46.
Moffat N.D., Lantz T.C., Fraser R.H., Olthof I. 2016. Recent vegetation change (1980–2013) in the tundra ecosystems of the Tuktoyaktuk coastlands, NWT, Canada. Arctic, Antarctic, and Alpine Research, 48: 581–597.
Nelson P.R., Roland C., Macander M.J., McCune B. 2013. Detecting continuous lichen abundance for mapping winter caribou forage at landscape spatial scales. Remote Sensing of Environment, 137: 43–54.
NIST/SEMATECH. 2020. e-Handbook of Statistical Methods. Available at: http://www.itl.nist.gov/div898/handbook/prc/section2/prc222.htm
Pedregosa F., Varoquaux G., Gramfort A., Michel V., Thirion B., Grisel O., 2011. Scikit-learn: machine learning in Python.’, Journal of Machine Learning Research, 12: 2825–2830.
Pilliod D.S., Arkle R.S. 2013. Performance of quantitative vegetation sampling methods across gradients of cover in great basin plant communities. Rangeland Ecology & Management, 66(6):634–647.
Riihimäki H., Luoto M., Heiskanen J. 2019. Estimating fractional cover of tundra vegetation at multiple scales using unmanned aerial systems and optical satellite data. Remote Sensing of Environment, 224: 119–132.
Rocha A.D., Groen T.A., Skidmore A.K., Willemen L. 2021. Role of sampling design when predicting spatially dependent ecological data with remote sensing. IEEE Transactions on Geoscience and Remote Sensing, 59(1): 663–674.
Rosso A., Neitlich P., Smith R.J. 2014. Non-destructive lichen biomass estimation in northwestern Alaska: a comparison of methods. PLoS ONE, 9(7): e103739.
Siewert M.B., Olofsson J. 2020. Scale-dependency of arctic ecosystem properties revealed by UAV. Environmental Research Letters, 15: 094030:1–14.
Symstad A.J., Wienk C.L., Thorstenson A.D. 2008. Precision, repeatability, and efficiency of two canopy-cover estimate methods in northern Great Plains vegetation. Rangeland Ecology & Management, 61(4): 419–429.
Theau J., Duguay C.R. 2004a. Lichen mapping in the summer range of the George River caribou herd (northern Québec-Labrador) using Landsat TM imagery. Canadian Journal of Remote Sensing, 30(6): 867–881.
Theau J., Duguay C.R. 2004b. Mapping lichen changes in the summer range of the George River caribou herd (Québec-Labrador, Canada) using Landsat imagery (1976-1998). Rangifer, 24(1): 31–49.
Tremblay B., Lévesque E., Boudreau S. 2012. Recent expansion of erect shrubs in the low Arctic: evidence from eastern Nunavik. Environmental Research Letters, 7: 035501.
Tsuyuzaki S., Iwahana G., Saito K. 2018. Tundra fire alters vegetation patterns more than the resultant thermokarst. Polar Biology, 41: 753–761.
Darren Pouliot https://orcid.org/0000-0002-7831-0792 [email protected]
Landscape Science and Technology Division, Environment and Climate Change Canada, 1125 Colonel By Drive, Ottawa, ONK1A 0H3, Canada
Mao Mao
Landscape Science and Technology Division, Environment and Climate Change Canada, 1125 Colonel By Drive, Ottawa, ONK1A 0H3, Canada
Robert H. Fraser https://orcid.org/0000-0002-8055-4403
Canada Centre for Mapping and Earth Observation, Natural Resources Canada, 580 Booth Street, Ottawa, ONK1A 0E4, Canada
Blair Kennedy
Landscape Science and Technology Division, Environment and Climate Change Canada, 1125 Colonel By Drive, Ottawa, ONK1A 0H3, Canada
Sylvain G. Leblanc https://orcid.org/0000-0003-2456-7119
Canada Centre for Mapping and Earth Observation, Natural Resources Canada, 580 Booth Street, Ottawa, ONK1A 0E4, Canada
Liming He
Canada Centre for Mapping and Earth Observation, Natural Resources Canada, 580 Booth Street, Ottawa, ONK1A 0E4, Canada
Wenjun Chen
Canada Centre for Mapping and Earth Observation, Natural Resources Canada, 580 Booth Street, Ottawa, ONK1A 0E4, Canada
You have requested "on-the-fly" machine translation of selected content from our databases. This functionality is provided solely for your convenience and is in no way intended to replace human translation. Show full disclaimer
Neither ProQuest nor its licensors make any representations or warranties with respect to the translations. The translations are automatically generated "AS IS" and "AS AVAILABLE" and are not retained in our systems. PROQUEST AND ITS LICENSORS SPECIFICALLY DISCLAIM ANY AND ALL EXPRESS OR IMPLIED WARRANTIES, INCLUDING WITHOUT LIMITATION, ANY WARRANTIES FOR AVAILABILITY, ACCURACY, TIMELINESS, COMPLETENESS, NON-INFRINGMENT, MERCHANTABILITY OR FITNESS FOR A PARTICULAR PURPOSE. Your use of the translations is subject to all use restrictions contained in your Electronic Products License Agreement and by using the translation functionality you agree to forgo any and all claims against ProQuest or its licensors for your use of the translation functionality and any output derived there from. Hide full disclaimer
© 2022. This work is published under http://creativecommons.org/licenses/by/4.0/ (the “License”). Notwithstanding the ProQuest Terms and Conditions, you may use this content in accordance with the terms of the License.
Abstract
Effective plot-based field sampling involves a trade-off between implementation efficiency and sample error. Optimal field sampling therefore requires quantifying the sample error under various sampling designs. For remote sensing applications, it is also important to understand how field sample error and training sample size (the number of pixels) affect the retrieval of surface properties. In this research, drone imagery was used to simulate field plots and investigate plot sampling error for forage lichen cover in relation to plot size, number of plots, and sampling strategy. The effect of this error on remote sensing-based lichen cover retrieval was evaluated using varying training sampling sizes in two different study regions in northern Canada. Results showed that cover with high spatial variability increased the number of plots or plot size required to achieve a specified level of error. For lichen cover retrieval at moderate spatial resolution (10–30 m), field sampling (plot size and number of plots) did not have as significant of an effect as regional differences (spectral separability of cover types), sensor, and the number of pixels used for model training. This plot simulation approach using drone images can be applied to other surface properties and regions to provide field sampling guidance.
