Unified climate AI framework for time series prediction of river discharge, wildfires, and carbon fluxes across multiple ecosystems

Land-surface responses including river discharge, wildfires, and carbon exchange are often predicted with domain-specific models that are costly to maintain and perform poorly where observations are sparse. Unified deep learning models that share parameters across domains could borrow information from data-rich domains to improve generalization in data-poor ones, but remain underexplored. We tested this approach using datasets from (1) process-model simulations of gross primary productivity (GPP) and ecosystem respiration (RECO) in the circumpolar Arctic; (2) simulations of GPP, RECO, growth and maintenance respiration for North American rangelands; and (3) observed river discharge, active fire counts, and burned area in Amazon river basins. We pretrained a shared transformer encoder with domain-specific input and output layers across all three domains and fine-tuned it separately for each. At held-out sites, the unified model outperformed domain-specific baselines for most targets. Improvements were largest in the data-scarce Amazon, where median Kling–Gupta efficiency increased from 0.32 to 0.69 for discharge, 0.17 to 0.49 for fire counts, and −0.05 to 0.43 for burned area, primarily through better representation of long-term means and variability. Predictions also improved in the data-rich Arctic (GPP: 0.74 to 0.85; RECO: 0.61 to 0.72), through more stable optimization. Rangeland predictions were already strong, and only two of four targets improved with joint training. Pairwise ablations identified Arctic as the most beneficial training partner, whereas pairing rangeland with dissimilar Amazon was detrimental. Overall, multi-domain training improves land-surface predictions over domain-specific baselines and offers a mechanism for leveraging simulations to strengthen predictions in data-scarce ecosystems.

Authors

Institutions

Publication Details

Published
2026-11-13
DOI
https://doi.org/10.22541/essoar.15012181/v1
Primary Topic
Hydrological Forecasting Using AI
Type
preprint
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
preprint

Unified climate AI framework for time series prediction of river discharge, wildfires, and carbon fluxes across multiple ecosystems

Brendan M. Rogers, Andrea D. de Almeida Castanho, Elchin E. Jafarov, Andrew L. Mullen et al.
2 links
Hydrological Forecasting Using AI
preprint

Unified climate AI framework for time series prediction of river discharge, wildfires, and carbon fluxes across multiple ecosystems

Brendan M. Rogers, Andrea D. de Almeida Castanho, Elchin E. Jafarov, Andrew L. Mullen, José Lucas Safanelli, Dominick Dusseau, Sandeep Poudel, Yili Yang
preprint en

Abstract

Land-surface responses including river discharge, wildfires, and carbon exchange are often predicted with domain-specific models that are costly to maintain and perform poorly where observations are sparse. Unified deep learning models that share parameters across domains could borrow information from data-rich domains to improve generalization in data-poor ones, but remain underexplored. We tested this approach using datasets from (1) process-model simulations of gross primary productivity (GPP) and ecosystem respiration (RECO) in the circumpolar Arctic; (2) simulations of GPP, RECO, growth and maintenance respiration for North American rangelands; and (3) observed river discharge, active fire counts, and burned area in Amazon river basins. We pretrained a shared transformer encoder with domain-specific input and output layers across all three domains and fine-tuned it separately for each. At held-out sites, the unified model outperformed domain-specific baselines for most targets. Improvements were largest in the data-scarce Amazon, where median Kling–Gupta efficiency increased from 0.32 to 0.69 for discharge, 0.17 to 0.49 for fire counts, and −0.05 to 0.43 for burned area, primarily through better representation of long-term means and variability. Predictions also improved in the data-rich Arctic (GPP: 0.74 to 0.85; RECO: 0.61 to 0.72), through more stable optimization. Rangeland predictions were already strong, and only two of four targets improved with joint training. Pairwise ablations identified Arctic as the most beneficial training partner, whereas pairing rangeland with dissimilar Amazon was detrimental. Overall, multi-domain training improves land-surface predictions over domain-specific baselines and offers a mechanism for leveraging simulations to strengthen predictions in data-scarce ecosystems.

Woodwell Climate Research Center (US), Cornell University (US)
Climate action, Life on land
Hydrological Forecasting Using AI
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon