An uncertainty-aware spatial overlay framework for prioritizing conservation planning

Abstract Spatially explicit and uncertainty-aware frameworks for habitat-based conservation planning remain methodologically limited, particularly in data-scarce landscapes. This study developed a novel uncertainty-aware dual-filter spatial overlay framework that operationalizes ensemble inter-model disagreement as an active spatial filter, rather than a residual diagnostic, to delineate priority conservation zones where high habitat suitability and strong model agreement converge. Rheum ribes L., an endangered montane plant species in northeastern Iran, served as a representative case study. Four machine learning algorithms (RF, GBM, ANN, DNN) were calibrated using 11 environmental predictors, including Landsat-derived remote sensing indices and topographic, climatic, and soil variables. Individual predictions were integrated via a TSS-weighted ensemble, which achieved the highest predictive accuracy (AUC = 0.94, TSS = 0.75). Spatial congruence analysis (Schoener’s D, Hellinger’s I) further revealed that tree-based algorithms showed closer spatial alignment with the ensemble (GBM: D = 0.81, I = 0.89; RF: D = 0.73, I = 0.84) than neural network models, indicating systematic differences in niche representation across algorithm families. Spatial uncertainty quantification revealed that ensemble variance carries ecologically meaningful information, with high-suitability habitats consistently associated with low inter-model disagreement. Applying the dual-filter framework delineated 19.2% of the study area as priority habitat zones. The framework provides a reproducible and transferable approach for biodiversity-informed conservation planning across ecosystems and taxa.