Predicted Ecological States of the Semiarid Warm Sandy and Loamy Uplands Ecological Site Group in the Upper Colorado River Basin, Based on Vegetation Functional Group Cover Data from 2021

This dataset provides a predictive ecological state map for the Southwest Sandy and Loamy Uplands Ecological Site Group (ESG). We developed a machine learning approach to predict ecological states by training vegetation cover data from three remotely sensed datasets—Rangeland Analysis Platform (RAP), Rangeland Condition Monitoring Assessment and Projection (RCMAP), and Landscape Cover Analysis and Reporting Tools (LandCART) —using the quantified ecological states from the state-and-transition model (STM).

Dryland ecosystems, covering 45 percent of Earth's land and supporting over one-third of the global population, face significant threats from land degradation and ecological changes driven by climate change. Managing these ecosystems is complex, and science-based frameworks like Ecological Site Descriptions (ESDs) and STMs are essential tools guiding decisions to support ecological health and stakeholder values. However, the alignment of ESDs and STMs with small-scale soil survey maps limits their applicability to monitoring broader ecological processes. To close this gap, we extended these frameworks to larger landscapes by utilizing quantified ecological states from the data-driven STM for the Semiarid Warm Sandy and Loamy Uplands ESG.

We hypothesized that a global random forest model, aggregating data from all three datasets, would outperform models based on individual datasets by leveraging their collective strengths. The global model indeed outperformed individual dataset models, making the most accurate predictions overall and for the Invaded ecological state when evaluated with an independent testing dataset. In a case study of two neighboring fifth-level watersheds, we found low spatial agreement between the individual dataset models, with fewer than one-third of pixels showing the same predicted ecological state across all three models. However, the global model showed greater pixel-level agreement with each single-dataset model than those models did with each other, suggesting it integrated information from each to arrive at a more consistent prediction. We selected the global model as the best model and used it to generate a predictive map of ecological states across the entire Southwest Sandy and Loamy Uplands Ecological Site Group.

The ecological state maps generated through this approach translate the STMs across the entire ESG, providing a valuable tool for land management at both watershed and larger landscape scales. These maps can facilitate land condition assessments, support development of resource management plans, and help identify priority areas for restoration and conservation.