Appropriate Uses and Key Limitations

TerrAdapt:Cascadia is designed to provide a regional-scale perspective of our changing environment, helping users monitor and assess the status and trends of regional species and ecosystems, project potential future impacts of climate change, and prioritize landscapes for conservation actions to increase resilience to emerging threats. The following are examples of use-cases for which TerrAdapt:Cascadia was developed:

• Monitoring changes in our region's forests and rangelands and identifying drivers of losses (e.g. wildfire, natural resource extraction, and development) as well as gains (e.g., natural regeneration and restoration)

• Monitoring the status and trends of habitat and connectivity for species and ecosystems of the Cascadia region

• Projecting potential future climatic conditions, identifying climatic refugia, and assessing potential impacts of climate change on habitat and connectivity for regional species and ecosystems

• Integrating the dynamic monitoring, projections, and spatial priorities of TerrAdapt:Cascadia into an adaptive management program where cycles of decision-making are informed by assessments using up-to-date information.

• Communicating the need to conserve important habitats using the powerful storytelling capabilities of the TerrAdapt:Cascadia map and dashboard.

---

Though TerrAdapt:Cascadia offers powerful new capabilities to monitor, project, and prioritize, it is important to also understand it's limitations. Most of these stem from the limitations inherent in models and satellite imagery. Recent advances in cloud computing, AI, and remote sensing have greatly improved the accuracy and reliability of the models that TerrAdapt:Cascadia relies on, however, models are simplifications of real-world complexity and it is possible to use them inappropriately to drive decision-making and in communications if the scale and detail of the questions exceed the model's ability to answer reliably. Some key limitations of our models and data include:

• At the level of an individual pixel in our gridded raster datasets (typically 30m - 100m resolution), model predictions can be unreliable due to classification or prediction error inherent in models. For example, classifying landcover from satellite imagery is challenging when pixels contain a mix of different classes or the classes have similar reflectance characteristics. Conducting assessments at the level of small parcels can be unreliable compared to an assessment over a broader area, where error in a few pixels will not meaningfully affect the result of aggregating over a large number of mostly accurate pixels.

• Projecting future climate and impacts from climate change is inherently unreliable due to error in global climate models, inaccuracies in downscaling global models to regional scales, uncertainties in future emissions, natural climate variability, and many other factors. Thus, projections of potential climatic refugia, future suitable habitat, and climate connectivity areas should be viewed as windows into potential futures, not predictions of likely outcomes. Understanding the potential future risks from climate change can help raise awareness of vulnerabilities and potential actions that can be taken to mitigate risks. Assessing the spatial agreement among a range of model predictions using different assumptions about emissions may provide greater confidence in areas with strong agreement.

• Empirical models, such as many of our species habitat models, are highly sensitive to any biases in the field data used to train them. For example, if species occurrence data used to train a habitat model is biased towards animals selecting habitat in a specific way (e.g., due to local adaptations, sex or age specific behaviors, etc.), it may not be reliable when projected more broadly to other areas with different environmental conditions or to other animals exhibiting different habitat selection behaviors. We strive to use field data broadly representative of the range of environments and selection behaviors exhibited across the modeling region. However, models will never capture the full breadth of potential habitat use. As such, there will always be individuals using and moving through habitat in ways not well predicted by our models.