Author(s): Giovanni Negro; Mauro Carolli; Andrea Andreoli; David Faro; Guido Zolezzi; Stefano Fenoglio; Paolo Lo Conte; Paolo Vezza

Linked Author(s): Giovanni Negro, Mauro Carolli, Andrea Andreoli, Guido Zolezzi, Paolo Vezza

Keywords: Habitat modelling; MesoHABSIM; Random Forest; Geomorphic Units; Freshwater fishes

Abstract: Detecting patterns and relationships that are repeated or generalizable in nature is a main purpose for all ecological scientists. However, modelling such relationships do not usually ensure adequate transferability between similar contexts. In the framework of water resources planning and management, meso-scale habitat models demonstrated high potential to assess suitable environmental conditions for freshwater species. Particularly, the MesoHABSIM (MesoHABitat SImulation Model) approach (i) enables river habitat depiction in an effective manner, (ii) may include in the analysis a large range of biotic and abiotic environmental descriptors, and (iii) can make predictions based on sophisticated machine learning techniques. For the latter aspect, the use of Random Forest (RF) has recently increased to identify most important variables that are related to species presence and abundance. In the present study, the transferability capabilities of RF models were calibrated and validated in different streams across the Northern part of Italy. In particular, RF models for adult and juvenile brown trout (Salmo trutta), marble trout (Salmo marmoratus), Italian Vairone (Telestes muticellus), Italian barbel (Barbus plebejus), and bullhead (Cottus gobio) species were used to predict presence or absence of individuals within mesohabitats (or geomorphic units, GUs). Models were developed and calibrated with field data collected in reference sites of Piemonte and Valle d’Aosta regions, selected for their natural hydro-morphological conditions and habitat characteristics. Model transferability tests were performed in streams located in different regions and not included in the model calibration dataset. Predictive capacities of the models performed well in terms of accuracy (ranging from 75% to 82%) and true skill statistic (ranging from 0.52 to 0.75). In our opinion, the high predictive performances can be related to (i) the use of a proper spatial resolution of mesohabitat, recognized for its high ecological relevance, (ii) the adequate hydro-morphological characterization of the water courses, and (iii) the large number of environmental descriptors provided by the MesoHABSIM approach. Results showed that mesohabitat suitability criteria based on RF can be considered transferable among streams located in different regions of Northern Italy, especially when streams are characterized by kindred hydro-morphological and climatic conditions.

DOI: https://doi.org/10.3850/IAHR-39WC2521711920221044

Year: 2022