From big data to small scales: Machine learning enhances microclimate model predictions.

Area of Science:

• Ecology
• Environmental Science
• Remote Sensing

Background:

• Microclimates significantly influence organismal behavior, physiology, and distribution.
• Traditional physical heat-balance models for microhabitat temperatures often contain biases due to complex environmental factors and parameter uncertainties.
• These model limitations impede ecological research and conservation efforts, especially concerning climate change.

Purpose of the Study:

• To enhance the accuracy of microclimate models using high-resolution drone-based mapping and machine learning.
• To identify and correct systematic errors in physical heat-balance model predictions of ground temperatures.
• To provide a framework for more accurate microclimate estimations for ecological and conservation applications.

Main Methods:

• Utilized drone imagery to create detailed environmental maps (solar radiation, vegetation indices, skyview factor).
• Parameterized a physical heat-balance model with drone-derived data.
• Validated physical model predictions against drone-mounted infrared thermal maps.
• Applied a random forest machine learning model to correct prediction biases.

Main Results:

• Machine learning reduced mean absolute errors by over 30% and mean square errors by 50%.
• Prediction inaccuracies were consistently narrowed by the machine learning approach.
• Identified key bias drivers including vegetation cover, solar radiation, and height above ground.
• Drone-based approach demonstrated high applicability in open, sparsely vegetated habitats.

Conclusions:

• Machine learning effectively corrects biases in physical microclimate models, significantly improving prediction accuracy.
• The integration of drone-based remote sensing and machine learning offers a powerful tool for ecological studies and conservation.
• Findings support the development of climate-resilient management strategies by providing more reliable microclimate data.