Science at Earthshot

Applying AI, machine learning, and ecosystem physics modelling to global ecosystem restoration.

Research Snapshots

ARR Carbon Projections

Our platform for predicting restoration outcomes leverages (1) an expansive database of ecological outcomes worldwide; (2) the integration of ground observations with remote sensing through machine learning; (3) process understanding through land surface models and knowledge-guided machine learning, and (4) global scaling through automation.

Natural Risk Modeling

Fire, drought and disease threaten all ecosystems more than ever due to climate change. To quantify the likelihood of such natural risks to Earthshot projects, we use global models based on observations of historic impact, combined with Earth System Model projections of likely future climate change. Our approach helps accurately quantify the appropriate buffer pool, which directly improves project financing and quality.

REDD+ Scoping Automation in LandOS

Accurate deforestation and biomass baselines are essential for the development of robust REDD+ projects. The carbon verification process, however, can be long and costly. Our REDD+ Scoping automation allows for quick and simplified assessments that ensure project feasibility and quality.

Deforestation Risk Assessment Tool

Earthshot’s Deforestation Risk Tool uses multiple publicly available datasets based on satellite imagery to quantify land eligibility, deforestation baselines, and risk. Our approach generates a rigorous and detailed assessment that goes far beyond requirements for the most widely used greenhouse gas crediting programs. Integration with machine learning provides full automation, enabling us to evaluate global deforestation risk at scale.

Global Biomass History

We use space-based observations to improve our estimates of carbon stocks and biomass changes. Leveraging a suite of satellite observations, such as GEDI and Landsat, combined with a convolutional neural network, we track changes in aboveground biomass for every location on Earth, and every year.

Process Modeling

We use mathematical models that represent the complex process interactions within ecosystems, such as the Functionally Assembled Terrestrial Ecosystem Simulator (FATES). This allows us to predict intervention outcomes directly based on the processes that determine them. In FATES, trees of different species compete to determine the ultimate carbon uptake, biodiversity, water use, and biophysical feedbacks to the atmosphere.

Deep Forest: Visualizing Reforestation

Visualizing the negative consequences of climate change has a significant impact on people’s emotions, motivations, and behaviors. We do the opposite, allowing people to see the landscape our projects aspire to. Using new advances in generative adversarial networks, we directly visualize what an area could look like under reforestation. These real life images allow viewers to feel the impact of their actions, and are but the first step in leveraging AI to better connect people with the planet.

Open Science


Open source Python package to ingest data from database or similar data sources, perform necessary allometric scaling and transforms that prepare data, and generate carbon accumulation curves with the associated uncertainty.
“Now the world is recognizing the urgent need to do ecological restoration at scale, for hundreds of thousands of hectares in almost all ecosystems worldwide.

This is required by the Paris Agreements to meet our net zero goals, and requires a new way of thinking, and new science."
Dr Trevor Keenan
Chief Scientist
Read More

Science at Earthshot

Harnessing ecosystem models, decades of earth observations, and the power of AI, we guide projects to success despite ecological uncertainty. 

Nature Regeneration

Restoring a degraded landscape to native forest is a monumental task with uncertain outcomes. Future forest growth is dependent on complex interactions between species, the environment, and management practices.

And, due to climate change, fire, drought, and disease threaten ecosystems more than ever. Our unique modeling system enables us to guide the decisions necessary to restore natural systems despite these uncertainties and simultaneously fulfill community needs through agroforestry and improved forest management plans where livelihoods are dependent on stewardship of the local landscape.
Our systems fuse two approaches. In the first, we use machine learning to combine historical satellite observations with on-the-ground observations to infer future growth, risk, and project eligibility. In the second, our scientists integrate output from "The Community Land Model: Functionally Assembled Terrestrial Ecosystem Simulator" a computational model that simulates the dynamic interactions between vegetation, soil, and the atmosphere, to explore ecosystem feedbacks and predict responses to environmental changes. The combination of these two approaches enables our team to create informed predictions of forest growth that account for the interactions between climate, forest species, management decisions, and the local environment.

Open Science


Open source Python package to ingest data from database or similar data sources, perform necessary allometric scaling and transforms that prepare data, and generate carbon accumulation curves with the associated uncertainty.

Forest Conservation

A forest is more than trees.

Even in our high quality "Afforestation, Reforestation, and Revegetation" projects, the planted trees are but the scaffolding that supports restoration toward complex, natural ecosystems. From mosses to mahoganies, moths to monkeys, the different species in a forest play unique roles in nutrient cycling, soil formation, and water regulation, providing resilience to climate change and disturbances like fire and disease outbreaks. By conserving mature forests in REDD+ projects, we provide a refuge for thousands of native species that can then re-establish in restored forests, ultimately contributing to forest health, stability, and function across the landscape.
To build trust in REDD projects we adopt a jurisdictional approach to the allocation of avoided deforestation. This approach relies on a third-party quantification of both total deforestation and deforestation risk within a jurisdiction, such a country or province.  However, during the early stages of project development these quantities are modeled to assess the financial and ecological feasibility of the project. Additionally, accurate deforestation and biomass baselines for the types of forests being conserved are essential for the development of robust REDD+ projects at all stages. For these, we have developed cutting edge machine learning methods that incorporate time-series of global satellite data, manual interpretations of forest change, and biomass data from the literature to estimate the current state of the forest and the risk of deforestation at scale.

A Commitment to Science and Quality

At Earthshot Labs, we build projects that conserve and restore nature while building the knowledge and capacity for continual improvement. Our PhD scientists are collaborating with leading researchers in forest regeneration around the world to create a systematic review of outcomes from actively planted restoration  projects. Additionally, Earthshot works with and advises the Pucallpa Restoration Lab, an 11 ha experimental forest regeneration project in Peru designed to answer questions about species mixtures, planting density, plot clearing and fertilization regimes. We also strive to advance the state of the art in remote monitoring of biomass, growth, and landscape change through the development of our proprietary historical biomass models and our technological efforts toward running CLM FATES with high ecological resolution and downscaled climate data on off-the-shelf hardware.

In addition to our commitment to sharing our knowledge about restoration with other scientists, land managers, and fellow travelers on our journey to restore the earth, we are equally committed to promoting high-quality carbon standards like ABACUS and Verra's recent methodology for jurisdictional accounting of REDD+ projects.