Floating solar expansion and wildlife: Developing a machine-learning platform to monitor waterbird community dynamics
- Floating photovoltaic (FPV) systems have been expanding globally since 2007.
- Few studies have examined FPV- wildlife interactions.
- FPVs can alter aquatic conditions by reducing light, temperature, and oxygen.
- Broader ecological impacts, particularly on higher trophic levels such as waterbirds, remain poorly understood, creating challenges for sustainable FPV expansion.
Four non-mutually exclusive hypotheses that FPVs increase:
- (H1) overall waterbird abundance by
creating new water-locked surfaces - (H2) functional evenness by
supporting the co-occurrence of multiple feeding guilds such as wading and diving species - (H3) the frequency of maintenance
and terrestrial foraging behaviors - (H4) spatial clustering of waterbirds
on and around the array.
- n = 2 ponds (1 FPV, 1 control)
- n = 4 cameras / pond, spaced evenly
- 1 image / min from sunrise–sunset
- Monitoring: Aug 2024 → FPV deployed Oct 2025 → End Oct 2026
- Each image split into near (bottom) and far (top) zones
- Near zone – birds counted and identified to species
- Far zone – birds counted only
Pipeline for Biodiversity Monitoring
α = model intercept
A_near = total identified bird abundance (per camera per day)
A_far = total unidentified bird abundance (per camera per day)
R = number of species detected (per camera per day)
D = diversity index (Shannon H′ or Simpson 1−D, per camera per day)
E = sampling effort (usable frames or total minutes analyzed)
V = visibility or image quality score (exposure, glare, clarity)
T = treatment (FPV = 1, control = 0)
P = period (after = 1, before = 0)
DOY = day of year (used in smoother term for seasonal variation)
b_camera = random effect for camera ID (repeated measures)
b_date = random effect for shared daily conditions
ABUNDANCE MODELS:
near half:
log E[A_near] = α + β1 T + β2 P + β3 T×P + log E + γ1 V + s(DOY) + (1|camera) + (1|date)
far half:
log E[A_far] = α + β1 T + β2 P + β3 T×P + log E + γ1 V + s(DOY) + (1|camera) + (1|date)
RICHNESS MODEL:
log E[R] = α + β1 T + β2 P + β3 T×P + γ1 E + γ2 V + s(DOY) + (1|camera) + (1|date)
DIVERSITY MODEL:
E[D] = α + β1 T + β2 P + β3 T×P + γ1 E + γ2 V + s(DOY) + (1|camera) + (1|date)
NOTE:
A_near and A_far are modeled separately log E acts as the effort offset only in abundance models E and V enter as covariates in richness and diversity models