The recent proliferation of Earth observation (EO) data has facilitated tremendous growth in applications of this technology. EO data have become progressively more used in species modeling, due to easy access to data almost anywhere in the world and the ability to have repeat observations of the same area, i.e., monitoring [1
]. Additionally, EO data can provide continuous predictor variables which require no interpolation or sampling biases [1
]. These predictor variables are often the most useful in modern presence/absence species distribution models using techniques like Boosted Regression Trees [2
]. Wetland species, in particular, can benefit from habitat/species modeling with EO because these habitats are typically remote/inaccessible (boreal wetlands especially), and they are also very dynamic in both space and time. Wetland mapping and monitoring with EO sensors, such as optical, Synthetic Aperture Radar (SAR), and Light Detection and Ranging have been well studied [3
], but the application of these technologies to characterize species habitat within wetlands is the next logical step which can help habitat monitoring and conservation efforts.
The species of interest for this study is the Yellow Rail (YERA–Coturnicops noveboracensis
), which is an obligate wetland species that are typically very secretive. The majority of its known habitat is in the boreal forest zone [8
], and it is thought to have a small and declining global population (ca. 10,000–25,000 individuals [9
]). However, the species is also listed as data deficient because of the challenges associated with sampling and mapping the habitat this species requires. This led to it being listed as a species of special concern under Canada’s Species at Risk Act
. Most descriptions characterize its breeding habitat as graminoid wetlands with shallow standing water, and an understory layer of senescent vegetation [10
]. The Yellow Rail’s suitable habitat is expected to be highly specific, dynamic, and sensitive to human hydrological disturbance making it an ideal test case for EO species habitat modeling/monitoring applications.
Most previous studies using EO data to study wetlands have focused on wetland mapping and monitoring, rather than directly modeling wildlife habitat. The mapping of wetland areas versus other landcover can typically be accomplished with either optical data [12
] to identify wetland vegetation or with a Digital Elevation Model to identify topographically wet areas [5
]. The seasonal and dynamic nature of wetlands is typically captured with SAR backscatter, due to its ability to see through clouds and its strong ability to detect surface water/flooded vegetation [3
]. This hybrid data approach can typically achieve very high accuracies for wetland, wetland class, and landcover maps [5
], but, in the end, this categorical data may not be very useful for characterizing meaningful habitat for wetland species. Using these EO sources directly to model the species’ habitat may be a better method [16
]. This method can directly correlate a species’ habitat to EO data without the intermediate step of running a species model on top of the landcover classification models, which may compound errors. These categorical landcover classifications are often not updated regularly, making species monitoring difficult when using such approaches [1
]. Further, EO data can be specifically chosen to detect habitat features of interest to a given species. In the case of YERA, we would want EO data to detect and monitor the changing hydrology and capture the specific vegetation structure rather than a coarser wetland type designation.
SAR, and specifically polarimetric SAR, including both fully polarimetric and compact polarimetric configurations, have shown a strong ability to detect water/vegetation interactions and characterize wetland vegetation structure [17
]. Certain decompositions from SAR, such as the Freeman–Durden decomposition [19
], can separate SAR backscatter into useful parameters for characterizing wetland habitat: Double bounce–typically associated with flooded vegetation in wetlands; volume scattering–higher values associated with higher biomass, and rough surface–rougher surface at the scale of the wavelength. Moderate-resolution optical data should also theoretically detect graminoid vegetation, but may be limited for characterizing vegetation/water interactions. A DEM may not be of interest for YERA modeling as DEMs are typically used to distinguish wetland from upland areas, and we would expect habitat to only occur in flat wetland areas.
While species habitat modeling with EO is less common than landcover mapping with EO, there have been numerous studies that use EO for species conservation purposes. Prior studies have used optical data for species distribution models of the bird’s breeding habitat [20
] or used SAR/LiDAR to model specific habitat vegetation structures for bird species [23
]. Other studies have used SAR to map and monitor Panda habitat [25
], and to characterize wetland habitat in the Brazilian Pantanal [26
]. Similar polarimetric SAR and bird habitat studies have been done with Black Tern habitat in the Great Lakes region [27
]. In a related study to ours, [16
] used Sentinel-1, Sentinel-2 (multi-spectral optical), and DEM data to predict YERA species abundance. Species habitat monitoring is potentially a very powerful and under-utilized application of this technology, especially when monitoring dynamic habitats.
This study aims to build on the work from [16
] by investigating the importance of polarimetric SAR from RADARSAT-2 for developing more accurate predictive models of the YERA habitat. We also investigate yearly and seasonal habitat dynamics with RADARSAT-2 and Landsat-8 data to determine if we can track changes in habitat suitability in time. The objectives of this study are divided into four parts: (1) Investigate the relative importance of polarimetric SAR for predicting YERA wetland habitat, (2) characterize YERA habitat based on the EO variables (3) explore the ability of EO data to detect changes in YERA habitat year-to-year, and (4) assess the ability of polarimetric SAR and optical data to predict suitable YERA habitat across larger areas.
An initial exploratory analysis of the inputs in relation to occupancy of YERA showed noticeable differences between presence/absence sites in some of the inputs (Figure 2
). The rough surface component of the Freeman–Durden decomposition showed a large difference between occupied and unoccupied sites, with YERA preferring lower surface roughness (Figure 2
panel a). Most non-backscatter RADARSAT-2 inputs showed distinct differences between present/absent sites (Figure 2
panel b and c), while the Landsat-8 indices tended to overlap more between the two site types (Figure 2
shows the relative importance of each variable in the Boosted Regression Tree model. As expected, based on the figures above, the three most important variables are RADARSAT-2 inputs which control the majority of the model. RADARSAT-2 backscatter inputs (HH, HV, and VV) appear to have the lowest importance, while polarization ratios or indicators which incorporate phase information (Freeman–Durden components) appear to be the most important for YERA habitat prediction. Some Landsat-8 indices and bands are seen to be moderately important in predicting suitable habitats. The model with all inputs recorded an AUC value of 0.86, which is generally considered high. The model with RADARSAT-2 inputs only showed an identical AUC of 0.86 (rounding up), while the Landsat-8 only mode showed an AUC of 0.75.
Response of predicted YERA occupancy probability to all EO inputs was generated to assess ideal habitat. Figure 4
shows some of the more interesting/important response curves. CrossPolRatio (Figure 4
b) and DoubleBounce_FD (Figure 4
c) were generally positively correlated with occupancy probability. RoughSurface_FD (Figure 4
a) was strongly negatively associated with occupancy. VolumeScattering_FD (Figure 4
d) showed a more complex relationship with occupancy with probability peaking at low to moderate amounts of volume scattering. Landsat-8 inputs, such as B3 and B6 (Figure 4
e,f), showed higher predicted probability associated with higher reflectance values.
The two-year trend data can be seen in Figure 5
. All site types (loss, gain, no change) had the same positive/negative change between the years, but the slope of the change did differ between site types. Sites with a loss of YERA occupancy saw a stronger dip in the cross-pol ratio than sites with no change or gain. Double bounce values increased less for sites that gained YERA while sites that lost YERA were associated with lower NDVI slope.
Predicted YERA occupancy can be seen in Figure 6
. Most wetland areas were found to not be suitable YERA habitat (darker colors). The majority of high probability habitat is localized to a few small areas in the fen immediately west and east of McClelland Lake. Other prime habitat areas were small hexagons near the edges of lakes or streams. All treed wetland areas were shown to have near-zero YERA occupancy probability. Overall, this predicts YERA occupancy with an 82% accuracy when compared to the 30% test field sites. This prediction had a near-equal number of commission and omission errors, indicating a relatively useful prediction.