Author: Dominika Prajzlerová
Species Distribution Models (SDMs) are powerful tools that help scientists understand where species live—and why—by linking species occurrences to environmental conditions. These models often use remote sensing (RS) data from satellites to describe habitat conditions across large areas.
But one key question often goes overlooked: When is the best time to acquire this data?
To explore this, we conducted a study in the Czech Republic that shows just how much the seasonal timing of RS data can affect how well these models predict bird distributions.
How we tested it:
We analyzed 104 bird species using occurrence data from the Czech Breeding Bird Survey and Sentinel-2 satellite imagery collected monthly from March through September. We applied all commonly used approaches to incorporate RS data into SDMs—using imagery from individual months, full-season averages, seasonal extremes, and combinations of these time frames. For simplicity, we refer to these strategies as different “periods.”
What we found:
Timing matters—but its impact varies by habitat.
There was no universally best period for predicting bird distributions. Bird species in wetlands had the most stable and accurate predictions across periods. Forest species also showed consistent results, largely because their models relied heavily on vegetation structure, which remains constant throughout the season. However, these forest models were generally less accurate overall. In contrast, species inhabiting woodlands and human-modified environments showed greater variation in model performance across periods.
Predictor importance changes with the season.
The importance of different RS predictors shifted depending on the period, reflecting changes in vegetation growth and other habitat features. Surprisingly, NDVI barely worked in our models, despite its common use; instead, simpler measurements like green band reflectance often performed better.
Combining time periods helps but isn’t always the best: species-specific patterns are key.
Combining time periods often improves model performance, but it isn’t always the best approach—many species showed better results using data from a single period. While the stability of model performance across different periods is to some extent influenced by habitat, the impact of the period remains highly species-specific. This variability makes broad recommendations difficult and underscores the need to test multiple seasonal datasets to build reliable models.
🔗 Read the full paper here: [https://doi.org/10.1002/ecog.07935]
Figure. Model performance (AUC) and predictor importance for the 3 most stable and 3 most unstable bird species across different seasonal periods. The x-axis represents the periods, while the y-axis shows AUC values and predictor importance scores, illustrating how stability and key predictors vary seasonally among species.