Dominant-feature Identification in Data from Gaussian Processes Applied to Finnish Forest Inventory Records

Conventional geostatistical methods often assume a single process across spatial scales, potentially masking scale-dependent patterns that originate from distinct underlying processes. Particularly, nearby locations exhibit similar values and thereby form connected structures - features - that vary across scales. While scale-space analysis aims to disentangle such overlapping structures and reveal scale-dependent features, there is no method available to detect statistically credible features in geostatistical data. Here, we introduce a scale-space decomposition method for identifying features in Gaussian process-modeled geostatistical data, which also enables the estimation of scale-dependent effects of predictor variables. Features are defined as statistically credible, scale-dependent structures identified by significant deviations from zero between differences of successive smooths of the data. To demonstrate these capabilities, we applied the approach to Finnish forest inventory data from the 1920s. We identified two essential spatial scales in basal area of common tree species: plot-to-plot variation and regional scale. Our scale-dependent analysis reveals that edaphic factors consistently influence all species across scales, while anthropogenic drivers show contrasting scale-specific effects: slash-and-burn agriculture negatively affects spruce at both scales but shows opposite effects on birch at different scales. These insights advance the understanding of historical forest ecology and demonstrate the utility of our approach.