1 of 18

Upscaling Wetland Fluxes using AI

Ashley Brereton, Zelalem Mekonnen, Bhavna Arora, Bill Riley, Kunxiaojia Yuan, Yi Xu, Yu Zhang, Qing Zhu, Tyler Anthony, Adina Paytan

2 of 18

Introduction

Wetlands are low oxygen environments

This slows down the decomposition of organic matter

Good for keeping carbon locked in the soil

Bad for Methane release

Goal: Complicated and nonlinear system – get the computer to work it all out.

3 of 18

Typical wetland site in the Delta

4 of 18

Region of interest: �Sacramento-San Joaquim Delta

Site Code	Site Name	Water Type	Salinity	Years of Data (Full)	Start Date
US-Myb	Mayberry Wetland	Non-Tidal	Fresh	13	2010
US-Tw1	Twitchell Wetland West Pond	Non-Tidal	Fresh	12	2011
US-Tw4	Twitchell Island East End Wetland	Non-Tidal	Fresh	10	2013

The Delta has a disproportionately large volume of CO₂and methane flux data (35 site years)

Home to an expansive area of wetlands (1000km²)

Perfect test case for model upscaling application

5 of 18

Model Training: Schematic

Target

AI model in training

Features (predictors)

6 of 18

Model Upscaling: Schematic

Model prediction

Features

Trained AI model

7 of 18

Model targets

CO₂flux

Methane flux

Target 1:

Target 2:

8 of 18

Model features (predictors)

Application: Regional upscaling

Important consideration: Data availability at the regional scale

Gold standard in-situ measurements are not available at the regional scale

WLDAS (1km, daily)

Weather

Air temp
Solar
Wind

Subsurface

Soil temp
Soil moisture
Water table depth

LANDSAT (30m, weekly)

Vegetation

Cover
Color

9 of 18

Model suite

Many studies use fancy models, but leave out simpler models, such as linear regression

The motivation for using fancy models is that the extra complexity is justified due to superior performance over simpler models – so this must be demonstrated!

So we establish a model framework:

Training a group of models

From linear regression (simple) to neural networks (complex)

… and in between (decision trees, support vector machines, gradient boosting)

Model Name	Category	Description	Key Strengths
Linear Regression	Regression	Fits a linear relationship between predictors and fluxes	Simple baseline, easily interpretable
Random Forest	Ensemble of Decision Trees	Aggregates multiple decision trees to enhance prediction stability	Robust to nonlinearity, reduces overfitting
Support Vector Machine (SVM)	Kernel-Based Method	Uses flexible kernels to find optimal separating hyperplanes	Effective in high dimensions, adaptable kernels
LightGBM	Gradient Boosting	Employs iterative boosting with efficient tree growth	Fast, memory-efficient, handles large datasets
XGBoost	Gradient Boosting	Improves boosting with regularization and efficient computations	Manages outliers, handles sparse data well
LSTM Neural Network	Recurrent Neural Network	Captures temporal dependencies in sequential data inputs	Ideal for time-series, learns long-term patterns
GRU Neural Network	Recurrent Neural Network	Similar to LSTM but streamlined with fewer parameters	Efficient temporal modeling, lower complexity

10 of 18

Model Training: Leave-One-Site-Out (LOSO)

Upscaling: learn from monitored sites and apply model to unmonitored sites

We should train our model appropriately in the same way

Training Recipe:

Take all site locations
Remove one site
Train model on remaining sites
Predict removed site
Cycle through all sites

Result:

‘Blind’ predictions of CO₂and methane flux for all sites.

Used to compare to observations to evaluate model performance (R², r, RMSE)

11 of 18

Model Training: Leave-One-Site-Out (LOSO)

We run the full model suite in LOSO training mode and compare performance

‘Good’ performance all round (R²> 0.5)

Complex models perform best

But the difference between simple and complex is small

12 of 18

Trained Models

CO₂flux

Methane flux

Tules

Rice

R² = 0.73

R² = 0.53

R² = 0.51

R² = 0.57

13 of 18

Aim

Now that we have a framework to upscale C fluxes, we test it on 2 types of vegetation

Tules (Environmentally ‘friendly’ crops)
Rice (Agriculture)

Purpose is to

Directly analyze the effectiveness of land use strategies
Implications for the environment

14 of 18

Tules Upscaling

NECB = Carbon Sequestration

Mostly negative

Carbon goes into the ground – Good!

RF = Radiative Forcing

Mostly negative, but variability exists

Greenhouse gas goes into the ground – Good!

15 of 18

Rice Upscaling

NECB = Carbon Sequestration

Mostly negative

Carbon goes into the ground – Good!

RF = Radiative Forcing

Mostly positive, but variability exists

Greenhouse gas goes into the atmosphere – Bad.

NOTE – Even the bad is better than the peat oxidation baseline

16 of 18

Comparing Regional Flux Predictions

TULES

RICE

NECB = -0.35 kg GHG = -0.2 kg

NECB = -0.25 kg GHG = +0.1 kg

17 of 18

Summary

Developed a robust Framework for upscaling C Fluxes using AI methods

Assessed two types of vegetation

Tules
Rice

Model supports wetland restoration

Model implies climate impact for land reclamation for agriculture

But model also indicates rice as a carbon sink despite it being a GHG source.

Future work – other agriculture practices in the Delta (corn, alfalfa) but more important tidal wetlands which are expected to be more complex.

18 of 18

Feed-forward feature selection

Target Variable	Step	Chosen Feature	R²	Target Variable	Step	Chosen Feature	R²
FCO2 TULES	1	Soil Adjusted Vegetation Index (SAVI)	0.59	FCO2 TULES	1	GNDVI (Greenness) NDVI	0.52
	2	Upward Sensible Heat Flux	0.73		2	Shortwave radiation	0.56
FCH4 TULES	1	Canopy Temperature	0.48	FCH4 RICE	1	Soil Adjusted Vegetation Index (SAVI)	0.34
	2	Soil Temperature	0.52		2	Emmisivity std	0.52
	3	GNDVI (Greenness) NDVI	0.53