Habitat-partitioning improves regional distribution models in multi-habitat species: a case study with the European bilberry

Abstract

Modelling the spatial distribution of multi-habitat species is challenging since they show multi-dimensional environmental responses that may vary sharply through habitats. Hence, for these species, the achievement of realistic models useful in conservation planning may depend on the appropriate consideration of habitat information in model calibration. We aimed to evaluate the role of different types of habitat predictors, along with habitat-partitioning, to improve model inference, detect non-stationary responses across habitats and simulate the impact of sampling bias on spatial predictions. As a case study, we modelled the occurrence of the multi-habitat plant species bilberry (Vaccinium myrtillus) in the Cantabrian Mountains (NW Spain), where it represents a basic trophic resource for threatened brown bear and capercaillie.We used MaxEnt to compare a baseline model approach calibrated with topo-climatic variables against three alternative approaches using explicit habitat information based on vegetation maps and remote sensing data. For each approach, we ran non-partitioned (all habitats together) and habitat-partitioned models (one per habitat) and evaluated model performance, overfitting and extrapolation. The highest performance was for habitat-partitioned models including habitat predictors. The lowest overfitting was for the baseline non-partitioned model, at the cost of achieving the highest predicted fractional area. The extrapolation success of habitat-partitioned models was low, with the highest performance for the baseline approach. Our results highlight that multihabitat species responses are non-stationary across habitats, with habitat-biased data resulting in weak spatial predictions. When modelling the distribution of multi-habitat species at regional scale, we recommend using habitat-partitioned models including habitat predictors, either vegetation maps or remote sensing data, to improve the realism of spatial outputs and its applicability in regional conservation planning.3

One-Class Classification of Natural Vegetation Using Remote Sensing: A Review

“…At the plant-community level, the typology studied ranged from plant associations [49,50] to natural habitats [5,[51][52][53]. Interestingly, Bradter et al [49] and Fenske et al [50] indicated that increasing the vegetation hierarchical level (i.e., from habitat to plant association or species) decreases OCC accuracy, while Suárez-Seoane et al [54] demonstrated the opposite. In addition, Suárez-Seoane et al [54] and Connor et al [55] showed that vegetation classes with a narrow ecological niche had higher classification accuracy than that with a wider niche.…”

“…Interestingly, Bradter et al [49] and Fenske et al [50] indicated that increasing the vegetation hierarchical level (i.e., from habitat to plant association or species) decreases OCC accuracy, while Suárez-Seoane et al [54] demonstrated the opposite. In addition, Suárez-Seoane et al [54] and Connor et al [55] showed that vegetation classes with a narrow ecological niche had higher classification accuracy than that with a wider niche. To improve the classification accuracy of plant communities, Tang et al [56] developed an approach that groups species into spectrally discriminating phenological groups based on a k-means classification applied to PCA axes of MODIS spectral variables.…”

“…The vegetation variable (i.e., spectral values) is used most often (n = 99) and comes exclusively from RS data. Specifically, several studies showed the utility of variables that characterize vegetation phenology derived from multispectral satellite data to map natural vegetation, such as MODIS [82,83], Landsat [54], ASTER [84], RapidEye [51], or hyperspectral airborne data [85]. Although Sentinel-2 data are increasingly relevant for OCCs, their use remained limited to a few acquisitions [45,[71][72][73].…”

“…Conversely, many studies have reported not only that adding a spectral variable to environmental variables increases OCC accuracy [82,[110][111][112], particularly when classifying climate-unstructured plant species [59], but also that the spectral variable often contributes the most [36,46]. Two other studies reported that adding spectral variables did enhance the spatial scale of the map substantially [54,106]. Only 25 studies used at least four variable types.…”

“…Regarding reference data, their quality is also an issue for OCC performance [54]. Although new reference plots can be collected according to a specific protocol suitable for classifying actual area of vegetation classes at local or regional scales, using field reference databases is inevitable when classifying larger sites and/or performing temporal monitoring.…”