Skip to content
Ecological Intelligence

Ecological evidence.
Independent perspective.
Explore Wild Lands ↗

Ecological Intelligence / Subject

machine learning

Peer-reviewed North America

Wetland carbon models scaled CO₂ better than methane

Evidence
Peer-reviewed
Geography
Southeastern United States
System
Freshwater and salt-marsh wetlands

WHAT THE EVIDENCE SAYS

Using data from 24 freshwater and salt-marsh sites, researchers compared machine-learning models for gross primary production, ecosystem respiration, net ecosystem exchange and methane flux. CO₂-related performance was moderate, while methane models were substantially weaker, reflecting the strong local influence of soil and hydrologic conditions.

WHY IT MATTERS / OUR READ

Regional upscaling is valuable, but greenhouse-gas certainty is pool-specific. A model that performs acceptably for CO₂ exchange should not inherit the same confidence for methane.

CAVEATS / UNCERTAINTY

The site network and input variables constrain generalization, and predictive skill for methane was modest. Model agreement is not a substitute for field validation when management or accounting depends on site-level fluxes.

SOURCE

Ecological Processes — primary modeling study
View original evidence ↗

An ensemble model built from 24 southeastern U.S. wetland sites produced useful regional CO₂ estimates, but methane remained much harder to predict.

Read the Wild Lands assessment →
Peer-reviewed North America

Machine learning can fill gaps in wetland water-table records—but skill drops when yesterday’s water table is missing

Evidence
Peer-reviewed
Geography
Santee Experimental Forest, South Carolina, USA
System
Forested wetlands / coastal plain watersheds

WHAT THE EVIDENCE SAYS

Researchers used observations from 13 wells in three forested watersheds with Daymet climate data and GLDAS land-surface variables to reconstruct continuous daily water-table depth. Models that included recent water-table state achieved very high within-record performance; when that lagged observation was unavailable, reconstruction skill declined, although models such as XGBoost still provided useful daily estimates from broader environmental inputs.

WHY IT MATTERS / OUR READ

Gap-filling can make irregular wetland monitoring more useful for trend and hydroperiod analysis, but the performance difference matters operationally: occasional field measurements still carry information the climate grids cannot fully replace.

CAVEATS / UNCERTAINTY

The study comes from one forested coastal-plain research system. Reconstructed values are model estimates, not observations, and strong within-system validation does not guarantee transfer to different soils, drainage networks or management regimes. Performance is highest when recent observed water-table state is available.

SOURCE

Wetlands — primary study
View original evidence ↗

Models at South Carolina’s Santee Experimental Forest reconstructed daily water-table depth from sparse observations and gridded climate data, with lagged hydrologic state carrying substantial predictive value.

Read the Wild Lands assessment →
Peer-reviewed Africa

Lake Tana’s water-hyacinth models fit extremely well—but prediction is not causation

Evidence
Peer-reviewed
Geography
Lake Tana, Ethiopia
System
Large tropical lake and invasive aquatic-plant system

WHAT THE EVIDENCE SAYS

A Scientific Reports study trained conventional and deep-learning models on long-term monitoring from 27 Lake Tana sites to predict water-hyacinth expansion. Random forest produced the strongest reported fit, and variable-importance analyses emphasized total nitrogen, total phosphorus, chlorophyll-a and pH among the influential predictors.

WHY IT MATTERS / OUR READ

The model may be useful for surveillance and hypothesis generation, but high predictive performance should not be confused with a nutrient-management experiment. The strongest operational use is to identify conditions and places for closer monitoring, then test which levers actually change invasion dynamics.

CAVEATS / UNCERTAINTY

Predictor importance in machine learning does not establish causal drivers, and very high internal model performance requires validation under genuinely independent future or external conditions. The study's monitoring design and local hydrology also limit transfer to other lakes.

SOURCE

Scientific Reports
View original evidence ↗

Machine-learning models linked two decades of monitoring to water-hyacinth expansion, highlighting nitrogen, phosphorus, chlorophyll-a and pH while leaving the causal management story unresolved.

Read the Wild Lands assessment →

Keep exploring

Follow the systems. Find the connections.

WildlifeLandWaterOceansClimateCitiesOne HealthGovernanceMethods / Monitoring
Browse the archive →