Novel insights into habitat suitability for two Unionida freshwater mussels, Castalia ambigua Lamarck, 1819 (Hyriidae) and Anodontites elongatus (Swainson, 1823) (Mycetopodidae), is presented on the basis of hydraulic variables linked with the riverbed in six 500 m reaches in an eastern Amazonian river basin. Within the reaches, there was strong habitat heterogeneity in hydrodynamics and substrate composition. In addition, we investigated stressors based on landscape modification that are associated with declines in mussel density. We measured hydraulic variables for each 500 m reach, and landscape stressors at two spatial scales (subcatchment and riparian buffer forest). We used the Random Forest algorithm, a tree-based model, to predict the hydraulic variables linked with habitat suitability for mussels, and to predict which landscape stressors were most associated with mussel density declines. Both mussel species were linked with low substrate heterogeneity and greater riverbed stability (low Froude and Reynolds numbers), especially at high flow (low stream power). Different sediment grain size preferences were observed between mussel species: Castalia ambigua was associated with medium sand, and Anodontites elongatus with medium and fine sand. Declines in mussel density were associated with modifications linked to urbanization at small scales (riparian buffer forest), especially with percent of and distance from rural settlements, distance to the nearest street, and road density. In summary, the high variance explained in both hydraulic and landscape models indicated high predictive power, suggesting that our findings may be extrapolated and used as a baseline to test hypotheses of habitat suitability in other Amazonian rivers for Castalia ambigua and Anodontites elongatus, and also for other freshwater mussel species. Our results highlight the urgent need for aquatic habitat conservation to maintain sheltered habitats during high flow as well as mitigate the effects of landscape modifications at the riparian buffer scale, both of which are important for maintaining dense mussel populations and habitat quality.

During low flow, between October and December 2018, we selected six sites in the Caeté River, eastern Amazon, Bragança, Pará, Brazil, known to have mussels. We established one 500 m reach per site covering all available habitats with different sediment characteristics and hydrodynamics (e.g. meanders, backwaters, and straight stretches). We classified mesohabitats in each reach according to substrate characteristics and water flow for a better habitat description. We identified six distinct mesohabitats that differed in terms of riverbed stability, substrate classification and grain size, and water flow. Along each mesohabitat, we established two equidistant 100 m transects. We placed twenty-five 1 m² plots along each transect, 300 plots per reach, by selecting a random position in such a way as to avoid clumping of plots and spatial autocorrelation.

Current velocity (m/s), measured at 1 cm above the river bed using a digital Flowatch meter (precision 0.01 m/s), and depth (m) measured using a metric stick, were obtained at the center of each plot and used to calculate the complex hydraulic variables (README Table 1). We sampled for mussels in each plot as the last step of the fieldwork to avoid bias during measurements of hydraulic variables. We sampled mussels by manually excavating the sediment to a depth of approximately 15 cm, and by conducting semi-quantitative timed searches, which provide better spatial coverage and is useful for rare mussel species. Searches were carried out by a 2- to 3- member crew until all mussels present at each plot were removed. All mussels were identified to species, their density quantified and returned to the sediment.

Landscape variables were analyzed at two spatial scales: subcatchment scale (drainage area of 5 km² associated with each sampling site) and riparian buffer scale (200 m riparian buffer extending 1 km along the river margin from each sampling site). We extracted a set of thirteen landscape variables (README Table 2) including information on roads, urban settlements and riparian vegetation from both sides of the river (n = 12 at the site scale) using layers from OpenStreetMap (https://www.openstreetmap.org) available via the OpenLayers plugin in QGIS 3.12, and land use information from Instituto Brasileiro de Geografia e Estatística (IBGE) available at https://www.ibge.gov.br/geociencias/informacoes-ambientais/cobertura-e-uso-da-terra. Landscape variables were classified according to three main groups of stressors: riparian forest cover, land use and urbanization.