Climate-Aware Machine Learning for Above-Ground Biomass Estimation - Study compares temporal and spatial modeling

Aug 5, 2026

The study examines how data science, machine learning, and AI can address environmental challenges by estimating above-ground biomass (AGBM) from satellite imagery. It compares temporal and spatial modelling approaches for AGBM estimation and evaluates the AI-Climate Alignment Framework by Kaack et al. (2022) for supporting environmentally responsible model development. Tree-based and neural learners are trained on a small dataset using temporal and spatial representations, and results assess both estimation performance and carbon emissions.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages14

Reading Time35 min

AccessFree to download

Introduction

Motivation for climate-relevant biomass estimation

Literature Review

Remote sensing data and modelling paradigms

Temporal vs spatial modelling for biomass

Climate-Aware Machine Learning

Performance vs emissions in model comparison

Climate-Aware Machine Learning for AboveGround Biomass Estimation Aske Meineche, IT University of Copenhagen This study explores the role of data science, machine learning, and artificial intelligence in addressing environmental challenges, specifically focusing on the estimation of AboveGround Biomass (AGBM) using satellite imagery. The research aims to compare the effectiveness of temporal and spatial modelling techniques in AGBM estimation and to assess the utility of the AI-Climate Alignment Framework proposed by Kaack et al. (2022) in guiding environmentally responsible model development. A tree-based learner and a neural learner are trained on a small dataset, using a temporal and a spatial representation. The results show that the tree-based learner emits less carbon in inference and outperforms the neural learner when training on small samples. Keywords: Machine learning, remote sensing, above-ground biomass estimation, earth observation, geospatial data science, climate-aware machine learning. 1. Introduction The environmental crises regarding global climate change and threats to the world’s ecosystems have been widely accepted by governmental bodies and international organizations as one of the most important global challenges on the agenda. Important documents, such as the sixth Assessment Report authored by the International Panel on Climate Change (Adler et al. , 2022) and the third official accounting of the state of the planetary boundaries (Richardson et al. , 2023) have recently underscored the severity of the threats to the fundamental conditions making our planet habitable by human civilization. Making science-based decisions in the face of these crises requires natural resource management systems (Lister et al. , 2020) as field measurements are infeasible at scale (Ghosh and Behra , 2018) . Many countries currently mandate national forest inventory programs (NFI) as this tool provides support in both climate and environmental questions (Lister et al. , 2020) . NFIs contain a variety of measurements regarding forests, among other above ground biomass (AGB) . This can support climate modelling as change in biomass strongly influences carbon sequestration (Jevšenak and Skudnik, 2021), as well as ecosystem response to climate change (NovoFernández et al. , 2019) . While field measurements have traditionally been used in many countries (Ghosh and Behra , 2018; Magnussen et al. , 2018), remote sensing is increasingly being utilized as both primary and auxiliary data (Magnussen et al. , 2018; Lister et al. , 2020) . Performing these complex function mappings between remotely sensed data from either UAVs or satellites has in part been aided by the advances in machine learning, which is playing a significant part in modern AGB-estimation (Liet al. , 2020; Ghosh and Behra , 2018; Tamimina et al. , 2022) . However, as tree-based machine learning algorithms as used in Li et al. (2020), Ghosh and Behra (2018), and Tamiminia et al. (2022) differ significantly in parameters, performance, complexity and training times from artificial neural nets as used in Vawda et al. (2024) . As the data volume of remotely sensed images of terrestrial areas is significant it is relevant to consider complexity and emissions training and inference, rather than solely being guided by accuracy (Strubell et al. , 2019) . In this paper two modelling frameworks for Above Ground Biomass are compared with regards to both accuracy and emissions, relying on the novel framework from (Kaack et al. , 2022) in guiding emission measurements. 2. Literature Review Machine learning tools applied to remote sensing data is a common combination to predict, classify, estimate or delineate factors of interest, whether it is algal blooms (Hill et al. , 2020), land use (Bastini et al. , 2023), wind turbines (Zhou et al. , 2019) , or monitoring of photovoltaic plants (Costa et al. , 2021) . However, remotely sensed data can be gathered from

FAQs

What is the document’s main goal in AGBM estimation?

To estimate above-ground biomass from satellite imagery while comparing temporal and spatial modelling techniques and assessing climate-responsible development using an alignment framework.

Which models are compared for AGBM estimation?

A tree-based learner and a neural learner are trained on a small dataset with temporal and spatial representations.

How does the study evaluate models beyond accuracy?

It considers carbon emissions during inference and training complexity, not only performance metrics such as RMSE, R², or accuracy.

Aug 5, 2026

Upload by Theodore