ECOCLIM

Motivation

‍Predicting how climate change impacts ecosystems requires high-precision climate data. While global circulation models (GCMs) operate on coarse grids, species and biological processes interact with microclimates at scales under 1 km. Key applications, such as tracking species migration, defining growing seasons, and modeling ecological niches, rely heavily on precise local temperature and precipitation data. Conventional downscaling often overlooks these micro-level nuances, resulting in inaccurate environmental assessments. By treating species distributions as ecological sensors, we can extract fine-grained climate signals to significantly improve projection accuracy.

Proposed Approach / Solution

‍ECOCLIM applies advanced deep learning architectures (including CNNs and diffusion models) to learn complex climate patterns. To ensure results are both statistically accurate and physically meaningful, the models incorporate physical constraints alongside coarse ERA5 data and high-resolution CHELSA baselines.

We further enrich these models by incorporating biodiversity data, which reflects microclimatic realities that are difficult to calculate through pure physics alone. Because this same biodiversity data is often studied alongside the resulting climate projections, we implement rigorous validation with independent datasets to prevent circular reasoning. Finally, we will scale this framework to a global model, maximizing its utility worldwide.

Impact

‍By bridging the spatial resolution gap between global climate models and local ecological realities, ECOCLIM will provide high-precision climate data essential for advancing species distribution and ecological niche modeling. Beyond fundamental ecology, these refined projections will support applied decision-making in conservation planning, climate-resilient agriculture, and disaster risk management to help protect vulnerable ecosystems and communities. The lasting legacy of the project will be a globally applicable, high-resolution downscaling framework that is freely accessible to researchers and practitioners via the CHELSA and EnviDat data portals. Built on open data sources like ERA5, CHELSA, and GBIF, this transparent and reproducible tool will foster widespread cross-disciplinary collaboration in climate impact science.

‍

Figure 1. Conceptual architecture of the physics-informed deep learning model. Coarse-resolution ERA5 data (0.25°) serves as input to a convolutional neural network (CNN), while 1 km CHELSA data, meteorological station records (white dots), vegetation cover (green), and species occurrences (yellow dots) serve as multi-target outputs. A physical conservation constraint (equation, bottom) enforces mass balance, such as preserving total precipitation from coarse to fine scales, while ecological inputs help the model resolve fine-scale microclimatic patterns at target weather stations.

‍

Figure 2. Conceptual framework showing species as microclimatic "sensors" in complex terrain. Standard downscaling models often introduce localized errors, displacing climate boundaries such as the 6°C isotherm or 200 mm isohyet to inaccurate elevations (dashed lines). By integrating species distributions with known environmental thresholds (e.g., treeline limits, arid adaptations, or snow dependency) alongside meteorological station data, ECOCLIM uses machine learning to correct these elevation biases and produce accurate fine-scale climate surfaces (solid lines).