Data integration with uneven spatiotemporal coverage: A point-process approach for dynamic species distribution models

Abstract

Species distribution models (SDMs) are widely used to predict where and when species occur, but the data available to fit them often resolve space and time unevenly. Designed surveys, such as line-transect distance-sampling surveys, provide structured spatial coverage and the detection information needed to estimate detectability, but logistical and funding constraints often limit them to short, infrequent periods. Opportunistic or citizen-science records are collected more frequently across the year, but from areas shaped by observer access and interest rather than survey design, and usually without comparable detection information. As a result, models built from either source alone tend to capture spatial structure or temporal dynamics well, but not both, limiting predictions for mobile species in dynamic environments. Here, we present a general framework, based on a log-Gaussian Cox process (LGCP) fitted using integrated nested Laplace approximations (INLA), that links multiple data sources to a shared spatiotemporal measure of species density through source-specific observation models: a detection function for survey transects, and spatial constraints reflecting the footprint of opportunistic effort. We apply the framework to cetacean line-transect surveys and whale-watching records. In a simulation with one month of survey data and twelve months of opportunistic data, the integrated model recovered spatiotemporal distribution patterns well (median correlation with the true pattern = 0.91) and estimated monthly abundance with a mean error of +10.7%, outperforming single-source models. In a case study of common dolphins (Delphinus delphis) off mainland Portugal, integration improved the spatial accuracy of predictions, although weak covariate effects limited the temporal variation the model could resolve. These results show that integration adds the most value when the two data sources resolve complementary aspects of distribution, in this case broad spatial structure from the survey and temporal replication from whale-watching, while temporal predictions remain limited when covariates carry weak seasonal signal. The framework is general and can be extended to other taxa, systems, and combinations of structured and opportunistic data by substituting the relevant observation models.