Using Physics-Informed Computer Vision to Measure Site-Specific Wind Patterns Through Vegetation Motion Analysis
This research proposes a novel methodology for high-resolution wind field measurement in landscape contexts using vegetation as a distributed sensor network. By treating plants as spring-mass-damper systems responding to diurnal wind forcing, we can extract spatially-detailed microclimate data from monocular video timelapse sequences through optical flow analysis. The methodology integrates monocular depth estimation, human-in-the-loop camera geopose calibration, and prior information about vegetation structure to reconstruct 3D wind velocity fields. This approach enables evidence-based landscape design decisions regarding thermal comfort, species placement, and site organization without requiring dense meteorological instrumentation. The resulting data can inform computational fluid dynamics (CFD) validation, microclimate mapping, and performance-based landscape interventions.
Contemporary landscape architecture increasingly demands quantitative understanding of site microclimates to address thermal comfort, ecological performance, and climate adaptation. Wind patterns at the scale of individual sites—influenced by topography, vegetation structure, and built form—create significant spatial heterogeneity that affects both human experience and ecological function. However, traditional anemometry provides only point measurements, while computational fluid dynamics (CFD) simulations require extensive validation data and often fail to capture the complexity of vegetated landscapes.
Plants respond mechanically to wind forcing as flexible oscillators, with their motion encoding information about wind speed, direction, and turbulence characteristics. This mechanical response varies with diurnal cycles as wind patterns shift due to differential heating and topographic effects. By analyzing vegetation motion at high temporal and spatial resolution, we can reconstruct detailed wind fields that capture the microclimate complexity relevant to landscape design decisions.
Rather than installing expensive sensor arrays, this methodology transforms the existing vegetation into a dense, distributed measurement network. The approach is particularly valuable for: (1) rapid site assessment during the design process, (2) post-occupancy evaluation of landscape performance, and (3) validation of predictive models used in the design phase.
We model each observable vegetation element (branch, stem, or whole plant) as a damped harmonic oscillator subject to external wind forcing:
where:
By measuring the displacement x(t) and its derivatives through optical flow analysis, we can invert this relationship to estimate the wind forcing function. The natural frequency ω₀ = √(k/m) and damping ratio ζ = c/(2√(km)) can be identified from the vegetation's free oscillation characteristics (observed during wind lulls) or estimated from plant morphology and material properties.
Different vegetation elements respond to different frequency ranges in the wind spectrum. Leaves flutter at high frequencies (>1 Hz), small branches oscillate at intermediate frequencies (0.1-1 Hz), and whole trees sway at low frequencies (<0.1 Hz). By analyzing motion across these scales simultaneously, we can reconstruct both mean wind patterns and turbulent fluctuations.
Wind patterns exhibit strong diurnal cycles driven by differential solar heating. During daytime, convective mixing strengthens and wind speeds typically increase. In complex terrain, katabatic (downslope) and anabatic (upslope) flows reverse with the diurnal cycle. By capturing vegetation motion over 24-hour periods, we can characterize these temporal patterns and their spatial manifestation across the site.
Recent advances in deep learning enable robust monocular depth estimation from single images. We leverage models trained on diverse outdoor scenes (e.g., MiDaS, DPT) to generate per-pixel depth maps. While absolute scale remains ambiguous in monocular reconstruction, we address this through:
To project motion measurements into a site-referenced coordinate system, we require camera position and orientation (geopose). We employ a hybrid approach:
Knowledge of vegetation geometry at rest provides crucial constraints for the inverse problem. We incorporate: