publications | Martin Schwartz

2026

FORMSpoT: Revealing fine-scale forest disturbances from nation-wide 1.5 m forest canopy height time series

Martin Schwartz, Fajwel Fogel, Nikola Besic, Damien Robert, Louis Geist, Jean-Pierre Renaud, Jean-Matthieu Monnet, Clemens Mosig, Cédric Vega, Alexandre d’Aspremont, Loic Landrieu, and Philippe Ciais

Remote Sensing of Environment, Dec 2026

abstract In brief paper

Current large-scale satellite-based forest disturbance monitoring systems operate at 10–30 m resolution, which is too coarse to detect changes at the scale of individual trees resulting in a systematic underestimation of forest disturbances. Here, we introduce FORMSpoT (Forest Mapping with SPOT Time series), a decade-long (2014-2024), country-scale mapping of forest canopy height at 1.5 m resolution over France, together with FORMSpoT-Δ, annual disturbance polygons derived from height differences in the FORMSpoT time series. Canopy heights were derived from annual SPOT-6/7 composites using a hierarchical transformer model (PVTv2) trained on high-resolution airborne laser scanning (ALS) data. To enable robust change detection, we developed a post-processing pipeline combining co-registration and spatio-temporal total variation denoising. We find that (1) the French disturbance regime is dominated by small events. Sub-100 m2 disturbances alone represent 72% of all events, and disturbances below 0.1 ha account for 97% of events and 39% of the disturbed area. These events are largely missed by Sentinel-1/2 and Landsat-based products. (2) Validated against successive ALS revisits across 19 sites and 5087 NFI plot revisits, FORMSpoT-Δ provides reliable detection (F1 \textgreater 0.8) above 100 m2 while retaining sensitivity to finer events that coarser products do not capture. (3) At the national scale, FORMSpoT-Δ resolves contrasted disturbance regimes, from clear-cut-dominated dynamics in maritime pine plantations to diffuse, smaller disturbance events in mountain forests, and captures their temporal dynamics, including the salvage-logging signature of the 2017–2022 bark beetle crisis in northeastern France.

Today’s large-scale forest-monitoring systems work at 10 to 30 metres, too coarse to see individual trees, so they miss many small disturbances. My colleagues and I built FORMSpoT, a decade-long (2014 to 2024) map of French forest canopy height at 1.5 metre resolution from SPOT satellites, together with yearly maps of where trees were lost. This lets us detect forest disturbances tree by tree across a whole country.
Integrating global canopy height models with satellite data for improved forest inventory in Ukraine

Petr Lukeš, Viktor Myroniuk, Andrii Shamrai, Viktor Melnychenko, Martin Schwartz, and Jan Pauls

Agricultural and Forest Meteorology, Sep 2026

abstract In brief paper

Accurate forest monitoring in Ukraine faces unprecedented challenges due to ongoing Russian aggression and limited access to traditional inventory methods, requiring robust satellite-based alternatives for National Forest Inventory operations. This study develops a comprehensive framework integrating state-of-the-art canopy height models with registered multi-sensor satellite data to achieve airborne LiDAR-comparable accuracy in forest structural parameter estimation. Here, we evaluated six canopy height models (CHM) against in-situ National Forest Inventory (NFI) measurements of canopy height across Ukraine. The most recent European-scale CHM by Pauls et al. (2025) demonstrated superior performance with R² = 0.68 and RMSE = 4.65 m. Using this CHM as input alongside registered optical (Sentinel-2) and radar (Sentinel-1, ALOS PALSAR-2) observations, we developed a machine learning framework testing 20 algorithms to address whether satellite-derived canopy height enhances forest structural parameter estimation and whether country-trained models match regionally specific approaches. Feature importance analysis revealed CHM-derived height as the dominant predictor (45-60%), followed by Sentinel-2 shortwave infrared bands and vegetation indices. This approach achieved relative RMSE values of 8.8-12.3% across four forest structural parameters: growing stock volume (R² = 0.77), basal area (R² = 0.68), stand age (R² = 0.53), and diameter at breast height (R² = 0.48). Performance improvements of 49-144% were observed compared to models without height information. The framework successfully generates wall-to-wall forest inventory maps with robust transferability across diverse forest conditions. This research establishes satellite-based forest inventory as a viable alternative for data-scarce environments worldwide.

Ukraine’s forests cannot be surveyed on the ground during the war, so satellites are the only option. With Petr Lukeš and colleagues from CzechGlobe (Czech Republic), we combined global canopy-height models with multi-sensor satellite data to estimate forest structure in Ukraine with an accuracy close to airborne LiDAR. This offers a way to keep the national forest inventory running remotely.
ProtoTree: An Efficient and Generalizable Model for Individual Tree Point Clouds Analysis

Alvin Opler, Philippe Ciais, Ibrahim Fayad, Martin Schwartz, Gabriel Belouze, Sarah Brood, Alexandre D’Aspremont, Dimitri Gominski, Mathieu Aubry, and Loic Landrieu

IEEE Transactions on Geoscience and Remote Sensing, Jul 2026

abstract In brief paper

Accurate estimation of tree volume, biomass, and species is central to carbon accounting and biodiversity monitoring. However, current methods either depend on costly field inventories or heavy geometric fitting. We present ProtoTree, a prototype-based pretraining framework optimized to reconstruct a per-tree point cloud by deforming a single learnable prototype, yielding compact and interpretable geometric descriptors. These descriptors explicitly capture tree-scale structural variation and enable label-efficient regression of individual-tree wood volume, which is subsequently combined with species-specific wood density to estimate above-ground biomass (AGB) from raw TLS, ULS, or ALS point clouds. With only 180 labeled trees, ProtoTree reaches R^2 = 0.94 for volume prediction compared to R^2 = 0.76 for a fully supervised alternative. Combined with a supervised classifier to predict species, our pipeline reproduces TLS-derived AGB with an MAE of 0.20 ton (against 0.36 ton using standard allometry equations), while requiring }approx 3 orders of magnitude less computation than quantitative structure models (QSMs). These results highlight ProtoTree as a scalable, label-efficient framework for operational forest inventories. With robust segmentation, it could enable large-scale carbon stock estimation and ecological monitoring of single trees. Our code and models are accessible at https://github.com/LSCE-forest/ProtoTree

Measuring the volume, biomass and species of individual trees usually needs costly fieldwork or heavy 3D modelling. With Alvin Opler and colleagues, we developed PROTOTREE, an AI method that describes each tree by gently deforming a single learnable “template” shape, giving compact and interpretable descriptors that need very few labels to train.
Alarming decline in the carbon sink of European forests driven by disturbances

François Ritter, Philippe Ciais, Cornelius Senf, Maurizio Santoro, Yidi Xu, Agnès Pelissier-Tanon, Martin Schwartz, Ibrahim Fayad, Nuno Carvalhais, Martin Brandt, Rasmus Fensholt, Simon Besnard, and Valerio Avitabile

National Science Review, Jun 2026

abstract In brief paper

Forests are carbon sinks essential for climate change mitigation. However, increased harvests and natural disturbances across Europe have recently challenged this role. To project the future carbon sink capacity of Europe’s forests, we integrated country reports from the United Nations Framework Convention on Climate Change (UNFCCC) with remote sensing maps of disturbances and above-ground biomass. Our model simulates biomass dynamics at 18 km resolution from 2010 to 2030, predicting a 39% decrease in the EU-27 forest carbon sink, driven by disturbances outpacing biomass recovery. Consequently, the 2030 forest carbon sink will fall 27% short of the estimated target consistent with EU-27 climate objectives. We demonstrate that the three billion trees initiative is insufficient for climate change mitigation and needs to be combined with a 28% reduction in forest harvests from 2025 to 2030 to meet these targets.

With Francois Ritter and colleagues from LSCE, we combined official country reports with satellite maps of forest loss and biomass to project the future of Europe’s forest carbon sink. Our model predicts a 39% drop in the EU carbon sink by 2030, driven by harvesting and natural disturbances.
Sub-pixel mapping of disturbance and tree mortality dynamics from Sentinel-2 time series around the globe

Clemens Mosig, Teja Kattenborn, David Montero Loaiza, Janusch Vanja-Jehle, John Brandt, Nathan Jacobs, Subash Khanal, Eric Xing, Martin Schwartz, Helene C. Muller-Landau, Mirela Beloiu, Aurora Bozzini, Yan Cheng, Keenan Ganz, Björn Grüning, Henrik Hartmann, Jan Hempel, Stéphanie Horion, Samuli Junttila, Kirill Korznikov, Guido Kraemer, Milena Mönks, Davide Nardi, Paul Neumeier, Jonathan Schmid, Salim Soltani, Marie Therese-Schmehl, Josh Veitch-Michaelis, and Miguel Mahecha

Preprint, Feb 2026

abstract In brief paper

Elevated forest disturbances and excess tree mortality are increasingly reported worldwide. Yet existing assessments are either based on patchy terrestrial observations or on large-scale satellite products, which are limited in resolution to pixel-level, binary tree loss detection. This leaves a blind spot on fine-scale disturbances where only a few trees are declining in an otherwise intact canopy. Here, we present a methodology for annually mapping sub-pixel fractional cover of standing deadwood and trees from rolling four-year windows of Sentinel-2 time series. Fractional cover is the proportion of each pixel covered by dead or live tree crowns. To obtain globally distributed sub-pixel reference labels, we leveraged the crowd-sourced archive deadtrees.earth of centimeter-scale drone orthophotos with two globally calibrated semantic segmentation models to derive tree and standing deadwood masks, yielding 6.2 million labeled Sentinel-2 pixels. Spatial block cross-validation shows high performance for tree cover, Pearson’s r = 0.58–0.64, across biomes, and moderate performance for standing deadwood cover, Pearson’s r = 0.30–0.56. Viewed as a binary classification task, the model achieves a mean precision of 91% and recall of 90% for tree cover, and a precision of 54% and recall of 82% for standing deadwood. Validation against independently obtained and ground-validated data in all major biomes reproduced known disturbance patterns and timing at unprecedented spatial detail compared to state-of-the-art products. Our method provides the critically needed missing link between fine-scale round observations and low-resolution remote sensing products, allowing more realistic estimates of global trends in forest disturbance and tree mortality.

Most satellite products can only flag whole pixels as “forest lost”, missing the early stages where just a few trees are dying. With Clemens Mosig and colleagues from the University of Leipzig, we developed a method to estimate, within each Sentinel-2 pixel, the fraction covered by dead or living tree crowns, revealing fine-scale forest decline around the globe.
An outlook on the rapid decline of carbon sequestration and perspectives for an improved monitoring of French forests

Philippe Ciais, Chuanlong Zhou, Pascal Schneider, Martin Schwartz, Nikola Besic, Cédric Vega, and Jean-Daniel Bontemps

Comptes Rendus. Géoscience, Jan 2026

abstract In brief paper

In this study, we present and discuss changes in carbon storage in French forests from 1990 to 2022, derived from CITEPA statistics on forest carbon accounting. These statistics are primarily informed by National Forest Inventory (NFI) data collected from systematic samples of forest plots across Metropolitan France, as well as additional sources related to forest removals, soils or wood products. As NFI is designed to provide statistical estimations of forest growing stock, gains and losses only at the national or subnational levels but not to deliver detailed spatial outlooks on disturbances carbon losses from fires, droughts and insect attacks, we also outline a prospect for future improvements enabled by remote sensing and the development of multi-source inventories. At a national level, a continuing removal of CO2 from the atmosphere occurred from 1990 to 2022, as harvest and mortality-induced CO2 losses remained smaller than CO2 removals by forest growth and the increase in forest area (ca. 80 000 ha per year since 2005 but insignificant in terms of increased carbon stocks at present). The CO2 removal by forests was 49.3 MtCO2·yr−1 in 1990, increased to reach a peak of 74.1 MtCO2·yr−1 in 2008 and then quickly decreased down to 37.8 Mton CO2·yr−1 in 2022. The changes in CO2 removal by forests can be separated into three phases. From 1990 to 2013, the CO2 removal increased alongside the increasing growth of living trees. A spike in carbon loss was caused by the passage of the Lothar and Martin extra-tropical cyclones but forests recovered rapidly within a few years. In contrast, from 2013 to 2017, the CO2 removal by forests quickly decreased due to increasing CO2 losses from harvest and natural mortality and a trend of decreasing productivity (Hertzog L. R. et al., Sci. Total Environ. 967 (2025), article no. 178843), each process contributing almost equally. After 2017, the sink remained low and mortality rates stayed larger than during any of the previous years. The recent period is marked by climate shocks such as summer droughts and heatwaves in 2015, 2018, 2022, 2023. The full impacts of the droughts in 2022 and 2023 are not yet covered with full precision, as some of the sites measured by the national inventory before those droughts are still pending a second visit. Delayed tree mortality can also manifest years after a drought has occurred. At a regional level, contrasted trajectories were identified. Southern Mediterranean regions where forests have a low harvest rate have also experienced a lower increase in mortality and a sustained CO2 uptake. Despite high harvest intensities, the Landes plantations also show an increasing CO2 sink. In contrast, all northern regions and Corsica have seen a strong decline in their CO2 removal rates, except in the Ile-de-France region (larger Paris area), where the CO2 sink was constant during the last 30 years, possibly because many forests are used for recreation and are subjected to low harvest pressure. Two regions, the Hauts-de-France and Grand Est forests, stand out as becoming net emitters of CO2 to the atmosphere. Other regions where the CO2 sink declined and is now close to zero are Normandy, Corsica, and Bourgogne-Franche-Comté. A detailed analysis was conducted to identify where trees are dying in France, the regions with increased mortality, and which species and tree sizes are most affected. We conclude with a perspective on how traditional sample-based statistical estimation of forest carbon changes, as implemented in classical NFI approaches, can be complemented by highresolution satellite and LiDAR data, together with denser monitoring of mortality processes. Progress in remote sensing technologies supports both model-based approaches aimed at mapping the carbon budget and enhanced inventory techniques for accurate estimation at finer spatial scales. Given the limited continuity of some long-term forest flux estimates, we finally outline potential pathways to strengthen carbon sink quantification in the near future.

With Philippe Ciais and colleagues from LSCE, we tracked how the carbon stored in French forests changed from 1990 to 2022 using national inventory statistics, and showed that the forest carbon sink is declining. Because these statistics cannot pinpoint where carbon is lost to fires, droughts and insects, we set out how satellite remote sensing could fill that gap in the future.
Impact of UAV Flight Parameters and Acquisition Context for Canopy Height and Trunk Circumference Measurement Using LiDAR

Clément Battista, Frederic Frappart, Martin Schwartz, and Frédéric Baup

Preprint, Jan 2026

abstract In brief paper

UAV-borne LiDAR has emerged as a promising tool for tree-level forest inventory, yet the influence of flight parameters and acquisition context on the accuracy of structural metrics remains insufficiently documented. This study evaluates how UAV flight altitude, speed, area margins, number of returns, flight plan geometry, and seasonal conditions affect the estimation of canopy height and trunk circumference in a managed 0.23-ha poplar plantation in southwestern France. Twenty-two UAV flights were conducted between 2024 and 2025 using a DJI Matrice 300 equipped with a Zenmuse L1 LiDAR sensor, with systematically varied acquisition parameters. Tree-level metrics derived from point clouds were compared with field measurements, terrestrial LiDAR scans, airborne LiDAR (IGN LiDAR HD), and satellite-derived canopy height maps (FORMS-T).Canopy height estimation proved highly robust across UAV flights, showing close agreement with terrestrial LiDAR (R² = 0.99, RMSE = 0.33 m) and consistency with expected poplar growth rates (0.9–1.4 m·yr⁻¹). Reliable height retrieval required a minimum point density of approximately 1,000 pts·m⁻², and flight altitudes below about 60 m above the canopy. In contrast, trunk circumference estimation was more sensitive to acquisition parameters and environmental conditions. Accurate retrieval was achieved only under leaf-off conditions, with optimal performance obtained using low flight altitudes (≤ 80 m), triple-return acquisition, high margins, and flight paths aligned with plantation rows (R² = 0.76–0.90; RMSE = 12–35 cm, or 4–11 cm for DBH). Leaf-on acquisitions and understory regrowth substantially reduced accuracy. This study provides practical guidelines for optimizing UAV flight planning in plantation forests and improves the reproducibility of UAV-based forest structural monitoring.

Drones carrying LiDAR can measure individual trees, but the results depend on how the drone is flown. With Clément Battista and colleagues from CESBIO in Toulouse, we systematically tested how flight altitude, speed and other settings affect measurements of tree height and trunk size in a poplar plantation in south-western France. The findings help set good practice for drone-based forest surveys.
ECHOSAT: Estimating Canopy Height Over Space And Time

Jan Pauls, Karsten Schrödter, Sven Ligensa, Martin Schwartz, Berkant Turan, Max Zimmer, Sassan Saatchi, Sebastian Pokutta, Philippe Ciais, and Fabian Gieseke

Preprint, 2026

abstract In brief paper

Forest monitoring is critical for climate change mitigation. However, existing global tree height maps provide only static snapshots and do not capture temporal forest dynamics, which are essential for accurate carbon accounting. We introduce ECHOSAT, a global and temporally consistent tree height map at 10 m resolution spanning multiple years. To this end, we resort to multi-sensor satellite data to train a specialized vision transformer model, which performs pixel-level temporal regression. A self-supervised growth loss regularizes the predictions to follow growth curves that are in line with natural tree development, including gradual height increases over time, but also abrupt declines due to forest loss events such as fires. Our experimental evaluation shows that our model improves state-of-the-art accuracies in the context of single-year predictions. We also provide the first global-scale height map that accurately quantifies tree growth and disturbances over time. We expect ECHOSAT to advance global efforts in carbon monitoring and disturbance assessment. The maps can be accessed at https://github.com/ai4forest/echosat.

Global tree-height maps are usually single snapshots that cannot show change over time. With Jan Pauls and colleagues from the University of Münster, we introduced ECHOSAT, a global tree-height map at 10 metre resolution that stays consistent from year to year, using an AI model nudged to follow realistic tree growth. This makes it possible to track both growth and losses such as fires worldwide.

2025

Retrieving yearly forest growth from satellite data: A deep learning based approach

Martin Schwartz, Philippe Ciais, Ewan Sean, Aurélien De Truchis, Cédric Vega, Nikola Besic, Ibrahim Fayad, Jean-Pierre Wigneron, Sarah Brood, Agnès Pelissier-Tanon, Jan Pauls, Gabriel Belouze, and Yidi Xu

Remote Sensing of Environment, Dec 2025

abstract In brief paper

High-resolution mapping of forest attributes is crucial for ecosystem monitoring and carbon budget assessments. Recent advancements have leveraged satellite imagery and deep learning algorithms to generate high-resolution forest height maps. While these maps provide valuable snapshots of forest conditions, they lack the temporal resolution to estimate forest-related carbon fluxes or track annual changes. Few studies have produced annual forest height, volume, or biomass change maps validated at the forest stand level. To address this limitation, we developed a deep learning framework, coupling data from Sentinel-1 (S1), Sentinel-2 (S2) and from the Global Ecosystem Dynamics Investigation (GEDI) mission, to generate a time series of forest height, growing stock volume, and aboveground biomass at 10 to 30-m spatial resolution that we refer to as FORMS-T (FORest Multiple Satellite Time series). Unlike previous studies, we train our model on individual S2 scenes, rather than on growing season composites, to account for acquisition variability and improve generalization across years. We produced these maps for France over seven years (2018–2024) for height at 10 m resolution and further converted them to 30 m maps of growing stock volume and aboveground biomass using leaf type-specific allometric equations. Evaluation against the French National Forest Inventory (NFI) showed an average mean absolute error of 3.07 m for height (r2 = 0.68) across all years, 86 m3 ha-1 for volume and 65.1 Mg ha-1 for biomass. We further evaluated FORMS-T capacity to capture growth on a site where two successive airborne laser scanning (ALS) campaigns were available, showing a good agreement with ALS data when aggregating at coarser spatial resolution (r2 = 0.60, MAE = 0.27 m for the 2020–2022 growth of trees between 10 and 15 m in 5 km pixels). Additionally, we compared our results to the NFI-based wood volume production at regional level and obtained a good agreement with a MAE of 1.45 m3 ha-1 yr-1 and r2 of 0.59. We then leveraged our height change maps to derive species-specific growth curves and compared them to ground-based measurements, highlighting distinct growth dynamics and regional variations in forest management practices. Further development of such maps could contribute to the assessment of forest-related carbon stocks and fluxes, contributing to the formulation of a comprehensive carbon budget at the country scale, and supporting global efforts to mitigate climate change.

Existing satellite maps of forests are single snapshots, so they cannot show how a forest changes from one year to the next. My colleagues and I trained an AI model on Sentinel-1, Sentinel-2 and GEDI satellite data to produce yearly maps of the height, volume and biomass of France’s forests from 2018 to 2024, the first time forest growth has been tracked at the scale of individual stands from space alone. This opens the door to measuring how much carbon French forests absorb each year.
SERA-H: Beyond Native Sentinel Spatial Limits for High-Resolution Canopy Height Mapping

Thomas Boudras, Martin Schwartz, Rasmus Fensholt, Martin Brandt, Ibrahim Fayad, Jean-Pierre Wigneron, Gabriel Belouze, Fajwel Fogel, and Philippe Ciais

Preprint, Dec 2025

abstract In brief paper

High-resolution mapping of canopy height is essential for forest management and biodiversity monitoring. Although recent studies have led to the advent of deep learning methods using satellite imagery to predict height maps, these approaches often face a trade-off between data accessibility and spatial resolution. To overcome these limitations, we present SERA-H, an end-to-end model combining a super-resolution module (EDSR) and temporal attention encoding (UTAE). Trained under the supervision of high-density LiDAR data (ALS), our model generates 2.5 m resolution height maps from freely available Sentinel-1 and Sentinel-2 (10 m) time series data. Evaluated on an open-source benchmark dataset in France, SERA-H, with a MAE of 2.6 m and a coefficient of determination of 0.82, not only outperforms standard Sentinel-1/2 baselines but also achieves performance comparable to or better than methods relying on commercial very high-resolution imagery (SPOT-6/7, PlanetScope, Maxar). These results demonstrate that combining high-resolution supervision with the spatiotemporal information embedded in time series enables the reconstruction of details beyond the input sensors’ native resolution. SERA-H opens the possibility of freely mapping forests with high revisit frequency, achieving accuracy comparable to that of costly commercial imagery. The source code is available at https://github.com/ThomasBoudras/SERA-H#

Freely available satellites such as Sentinel are limited to about 10 metre detail, which can blur individual trees. With Thomas Boudras and colleagues from LSCE, we built SERA-H, a model that combines image super-resolution with time-series data to map canopy height in finer detail than the raw satellite images allow.
Satellite-based mapping of annual canopy height and aboveground biomass in African dense forests

Liang Wan, Philippe Ciais, Aurélien Truchis, Ewan Sean, Fabian Jörg Fischer, David Purnell, Gabriel Belouze, Ibrahim Fayad, Martin Schwartz, Yidi Xu, Yang Su, Maxime Réjou Méchain, Nicolas Barbier, Paul Tresson, Jean-François Bastin, Jan Bogaert, Arthur Vander Linden, Antoine Plumacker, Bhely Angoboy, Dieumerci Assumani, Thales Haulleville, Le Bienfaiteur Sagang, Laurent Durieux, Youngryel Ryu, Tackang Yang, Conan Vassily Obame, Thomas Bossy, Frédéric Frappart, Marc Peaucelle, Jean-Pierre Wigneron, Jerome Chave, Aida Cuni-Sanchez, Wannes Hubau, Hans Verbeeck, Pascal Boeckx, Jean-Remy Makana, Corneille Ewango, Elizabeth Kearsley, and Pierre Ploton

Frontiers in Remote Sensing, Nov 2025

abstract In brief paper

Accurate maps of canopy height (CH) and aboveground biomass (AGB) are needed for monitoring forests over large regions. Producing such data is particularly challenging over the complex, diverse and dense humid tropical forests of Africa where signal saturation observed from optical and radar satellites and complex responses in LiDAR data require advanced mapping techniques to capture high biomass and tall height values. Here, we trained a deep learning (U-Net) model to generate the first annual maps (2019–2022) of top CH at 10 m resolution over the African dense forest region, using Sentinel-1/-2 images trained on LiDAR-derived height data from the Global Ecosystem Dynamics Investigation mission (GEDI). To predict AGB from CH on a 30-m grid, we calibrated allometric models combining AGB data from field inventories, CH from our map, and wood density from a new high-resolution (1 km) map. The CH map has a mean absolute error (MAE) of 4.54 m and an underestimation bias of 1.54 m compared to independent airborne LiDAR data (5.93 m and 1.40 m compared to independent GEDI data). Evaluation of the AGB map against independent measurements from field sites suggests an improved accuracy (MAE = 79.65 Mg/ha, bias = 6.47 Mg/ha) compared to recent datasets such as ESA-CCI, NCEO, and GEDI L4B. Our map also captures the large-scale spatial gradients of AGB across African dense forests, as observed in a comprehensive dataset of forest concession measurements aggregated at a 1-km scale. Interpretable machine learning was used to assess the contribution of ancillary variables (e.g., climate, soil, forest type) to biomass prediction. While some variables were relevant, their inclusion failed to improve AGB estimates in high and low biomass extremes and introduced spatial artifacts, limiting their utility for consistent annual mapping. Together, our annual CH and AGB maps offer an open, scalable tool for monitoring forest disturbances and interannual biomass dynamics. Future work will focus on refining biomass–height relationships to further improve AGB estimation.

Mapping the tall, dense forests of tropical Africa from space is notoriously difficult because the satellite signals saturate. With Liang Wan and colleagues from LSCE, we trained a deep-learning model to produce the first annual maps (2019 to 2022) of canopy height and biomass for these forests at 10 metre resolution.
GEDI and Sentinel data integration for quantifying agroforestry tree height and stocks

Giovanni D’Amico, Elia Vangi, Martin Schwartz, Francesca Giannetti, Saverio Francini, Piermaria Corona, Walter Mattioli, and Gherardo Chirici

Journal of Environmental Management, Oct 2025

abstract In brief paper

Agroforestry is a strategic asset for combating climate change and mitigating the environmental impacts of agricultural intensification, offering a nature-based solution for enhancing landscape resilience. In particular, poplar plantations contribute to the development of ecological networks within homogeneous agricultural landscapes, while also producing high-demand plywood and sequestering CO2 in durable manufactured goods. Monitoring short-rotation poplar plantations requires frequent updates, which are infeasible with conventional National Forest Inventories (NFIs). Remote sensing (RS) has emerged as a highly effective tool for monitoring the structural variables of poplar plantations. This study aims to estimate the carbon stocks of poplar plantations in the Padan Plain, which spans approximately 46,000 km2 in northern Italy. To achieve this, we developed a 10m high-resolution canopy height model (CHM) using a deep learning U-Net approach, with Sentinel-1 and Sentinel2 multi-band imagery as predictors for GEDI waveforms derived tree height. The U-Net CHM for 2021 was evaluated with external validation data from NFI plots, achieving a mean absolute error of 2.6 m. Using annual poplar plantations data in the survey area, along with a poplar-specific yield table derived from terrestrial laser scanning, we applied the U-Net CHM to predict key forestry variables, including diameter at breast height (DBH), growing stock volume (GSV), aboveground biomass (AGB), and carbon stock (CS) in all stands. Results were compared with external validation data from NFI, yielding RMSE values of 30.7 %, 46.2 %, and 63.2 % for DBH, GSV, and AGB, respectively. Meanwhile, independent field surveys produced RMSE values of 19 % and 37.7 % for DBH and GSV, respectively. The average GSV estimated was 70 m3 ha 1, while total CS were 12 MgC ha 1. Based on poplar plantation maps for 2021 and 2022, we estimated the total harvested GSV of poplar trees to be 370,000 m3, equal to 66,000 MgC. The corresponding average harvested area was 1.5 ha, with an average yield of 130 m3 per hectare. The integration of multiple RS datasets with advanced machine learning techniques facilitates the effective monitoring of dynamic poplar plantations, for both mapping purposes and quantifying key forest variables relevant to climate change mitigation, such as carbon stocks.

Poplar plantations scattered across farmland store carbon and supply timber, but they change fast and are hard to keep track of. With Giovanni D’Amico and colleagues from the University of Florence, we combined GEDI LiDAR and Sentinel satellite data to estimate the height and wood stocks of these agroforestry trees in Italy.
A fused canopy height map of Italy (2004–2024) from spaceborne and airborne LiDAR, and Landsat via deep learning and Bayesian averaging

Yang Su, Nikola Besic, Xianglin Zhang, Yidi Xu, Saverio Francini, Giovanni D’Amico, Gherardo Chirici, Martin Schwartz, Ibrahim Fayad, Sarah Brood, Agnes Pellissier-tanon, Ke Yu, Haotian Chen, Songchao Chen, Alexandre d’Aspremont, and Philippe Ciais

Earth System Science Data Discussions, Sep 2025

abstract In brief paper

\textlessp\textgreater\textlessstrong class="journal-contentHeaderColor"\textgreaterAbstract.\textless/strong\textgreater Forests are vital for the carbon sequestration, biodiversity conservation, and climate regulation, making the precise and continuous monitoring of forest structure attributes such as canopy height essential. Here we present a two decades long (2004–2024), 30 m resolution annual canopy height dataset for Italy, developed using a time-series deep learning framework that integrates Landsat optical imagery with LiDAR observations. Two UNET models were independently trained using canopy height reference data from airborne laser scanning (ALS) and NASA’s Global Ecosystem Dynamics Investigation (GEDI) spaceborne LiDAR mission. Annual canopy height predictions from each model were fused using Bayesian Model Averaging (BMA) to enhance spatial consistency and temporal continuity. Validation against ground-based measurements from the Italian National Forest Inventory (NFI) demonstrated high predictive accuracy (mean absolute error = 3.98 m). To further evaluate the utility of our dataset, we derived a canopy height change-based disturbance product and validated it against observed events (mean precision = 0.64 for 2005–2016). In addition, we assessed post-disturbance recovery by monitoring canopy height regrowth in areas affected during 2004–2005, tracking changes annually through 2024 across various Italian biomes. Our results highlight the importance of integrating multi-source remote sensing data with deep learning and Bayesian data fusion for monitoring forest structural dynamics. The final dataset is publicly available via Zenodo and provides a reproducible and scalable resource to support forest research, ecological monitoring, and climate-related policy-making.\textless/p\textgreater

With Yang Su and colleagues from LSCE, we built a two-decade record (2004 to 2024) of forest canopy height across Italy at 30 metre resolution, blending Landsat satellite images with airborne and spaceborne LiDAR through deep learning and Bayesian averaging. Long time series like this show how forests have grown and changed over twenty years.
Open-canopy: Towards very high resolution forest monitoring

Fajwel Fogel, Yohann Perron, Nikola Besic, Laurent Saint-André, Agnès Pellissier-Tanon, Martin Schwartz, Thomas Boudras, Ibrahim Fayad, Alexandre d’Aspremont, Loic Landrieu, and Philippe Ciais

Proceedings of the IEEE/CVF conference on computer vision and pattern recognition (CVPR), Jun 2025

abstract In brief

Estimating canopy height and its changes at meter resolution from satellite imagery is a significant challenge in computer vision with critical environmental applications. However, the lack of open-access datasets at this resolution hinders the reproducibility and evaluation of models. We introduce Open-Canopy, the first open-access, country-scale benchmark for very high-resolution (1.5 m) canopy height estimation, covering over 87,000 km2 across France with 1.5 m resolution satellite imagery and aerial LiDAR data. Additionally, we present Open-Canopy-∆, a benchmark for canopy height change detection between images from different years at tree level—a challenging task for current computer vision models. We evaluate state-of-the-art architectures on these benchmarks, highlighting significant challenges and opportunities for improvement. Our datasets and code are publicly available at https://github.com/fajwel/open-canopy

Progress in mapping forests from satellites is slowed by the lack of shared, high-quality datasets to test methods on. With Fajwel Fogel and colleagues from the École Normale Supérieure, we released Open-Canopy, the first open, country-scale benchmark for very-high-resolution (1.5 m) tree-height mapping, covering more than 87,000 km2 of France with satellite images and aerial LiDAR. It also includes a benchmark for spotting where individual trees have changed from one year to the next, so teams can train and fairly compare their models.
DUNIA: Pixel-Sized Embeddings via Cross-Modal Alignment for Earth Observation Applications

Ibrahim Fayad, Max Zimmer, Martin Schwartz, Fabian Gieseke, Philippe Ciais, Gabriel Belouze, Sarah Brood, Aurélien de Truchis, and Alexandre d’Aspremont

Forty-second International Conference on Machine Learning (ICML), Jun 2025

abstract In brief paper

Significant efforts have been directed towards adapting self-supervised multimodal learning for Earth observation applications. However, most current methods produce coarse patch-sized embeddings, limiting their effectiveness and integration with other modalities like LiDAR. To close this gap, we present DUNIA, an approach to learn pixel-sized embeddings through cross-modal alignment between images and full-waveform LiDAR data. As the model is trained in a contrastive manner, the embeddings can be directly leveraged in the context of a variety of environmental monitoring tasks in a zero-shot setting. In our experiments, we demonstrate the effectiveness of the embeddings for seven such tasks: canopy height mapping, fractional canopy cover, land cover mapping, tree species identification, plant area index, crop type classification, and per-pixel waveform-based vertical structure mapping. The results show that the embeddings, along with zero-shot classifiers, often outperform specialized supervised models, even in low-data regimes. In the fine-tuning setting, we show strong performances near or better than the state-of-the-art on five out of six tasks.

With Ibrahim Fayad and colleagues from LSCE, we developed DUNIA, an AI method that learns a compact “fingerprint” for every pixel of a satellite image by aligning it with LiDAR data. Once learned, these fingerprints can tackle many environmental tasks, from mapping canopy height to classifying crops, often with very little extra training data.
Canopy height and biomass distribution across the forests of Iberian Peninsula

Yang Su, Martin Schwartz, Ibrahim Fayad, Mariano García, Miguel A. Zavala, Julián Tijerín-Triviño, Julen Astigarraga, Verónica Cruz-Alonso, Siyu Liu, Xianglin Zhang, Songchao Chen, François Ritter, Nikola Besic, Alexandre d’Aspremont, and Philippe Ciais

Scientific Data, Apr 2025

abstract In brief paper

Accurate mapping of vegetation canopy height and biomass distribution is essential for effective forest monitoring, climate change mitigation, and sustainable forestry. Here we present high-resolution remote sensing-based canopy height (10 m resolution) and above ground biomass (AGB, 50 m resolution) maps for the forests of the Iberian Peninsula from 2017 to 2021, using a deep learning framework that integrates Sentinel-1, Sentinel-2, and LiDAR data. Two UNET models were developed: one trained on Airborne Laser Scanning (ALS) data (MAE: 1.22 m), while another using Global Ecosystem Dynamics Investigation (GEDI) footprints (MAE: 3.24 m). External validation with 6,308 Spanish National Forest Inventory (NFI) plots (2017–2019) confirmed canopy height reliability, showing MAEs of 2–3 m in tree-covered areas. AGB estimates were obtained through Random Forest models that linked UNET derived height predictions to NFI AGB data, achieves an MAE of ~29 Mg/ha. The creation of high-resolution maps of canopy height and biomass across various forest landscapes in the Iberian Peninsula provides a valuable new tool for environmental researchers, policy makers, and forest management professionals, offering detailed insights that can inform conservation strategies, carbon sequestration efforts, and sustainable forest management practices.

With Yang Su and colleagues from LSCE, we mapped the height and above-ground biomass of the forests of the Iberian Peninsula (Spain and Portugal) at high resolution, using Sentinel-1, Sentinel-2 and LiDAR data with deep learning. The resulting maps help track carbon stocks and support forest management across the region.
State of the art and perspectives for remote sensing monitoring of carbon dynamics in African tropical forests

Thomas Bossy, Philippe Ciais, Solène Renaudineau, Liang Wan, Bertrand Ygorra, Elhadi Adam, Nicolas Barbier, Marijn Bauters, Nicolas Delbart, Frédéric Frappart, Tawanda Winmore Gara, Eliakim Hamunyela, IFO Suspense Averti, Gabriel Jaffrain, Philippe Maisongrande, Maurice Mugabowindekwe, Theodomir Mugiraneza, Cassandra Normandin, Conan Vassily Obame, Marc Peaucelle, Camille Pinet, Pierre Ploton, Le Bienfaiteur Sagang, Martin Schwartz, Valentine Sollier, Bonaventure Sonké, Paul Tresson, Aurélien De Truchis, An Vo Quang, and Jean-Pierre Wigneron

Frontiers in Remote Sensing, Feb 2025

abstract In brief paper

African tropical forests play a crucial role in global carbon dynamics, biodiversity conservation, and climate regulation, yet monitoring their structure, diversity, carbon stocks and changes remains challenging. Remote sensing techniques, including multi-spectral data, lidar-based canopy height and vertical structure detection, and radar interferometry, have significantly improved our ability to map forest composition, estimate height and biomass, and detect degradation and deforestation features at a finer scale. Machine learning approaches further enhance these capabilities by integrating multiple data sources to produce improved maps of forest attributes and track changes over time. Despite these advancements, uncertainties remain due to limited ground-truth validation, and the structural complexity and large spatial heterogeneity of African forests. Future developments in remote sensing should examine how multi-sensor integration of high-resolution data from instruments such as Planet, Tandem-X, SPOT and improved AI methods can refine forest composition, carbon storage and function maps, enhance large-scale monitoring of tree height and biomass dynamics, and improve forest degradation and deforestation detection down to tree level. These advancements will be essential for supporting science-based decision-making in forest conservation and climate mitigation.

African tropical forests are vital for the climate but very hard to monitor. With Thomas Bossy and colleagues from LSCE, we reviewed how satellites (optical, radar and LiDAR) can track their structure, carbon and loss, and set out what is still missing.
Remote-sensing-based forest canopy height mapping: some models are useful, but might they provide us with even more insights when combined?

Nikola Besic, Nicolas Picard, Cédric Vega, Jean-Daniel Bontemps, Lionel Hertzog, Jean-Pierre Renaud, Fajwel Fogel, Martin Schwartz, Agnès Pellissier-Tanon, Gabriel Destouet, Frédéric Mortier, Milena Planells-Rodriguez, and Philippe Ciais

Geoscientific Model Development, Jan 2025

abstract In brief paper

The development of high-resolution mapping models for forest attributes based on remote sensing data combined with machine or deep learning techniques has become a prominent topic in the field of forest observation and monitoring. This has resulted in the availability of multiple, sometimes conflicting, sources of information, but, at face value, it also makes it possible to learn about forest attribute uncertainty through the joint interpretation of multiple models. This article seeks to endorse the latter by utilizing the Bayesian model averaging approach to diagnose and interpret the differences between predictions from different models. The predictions in our case are forest canopy height estimations for metropolitan France arising from five different models. An independent reference dataset, containing four different definitions of forest height (dominant, mean, maximum, and Lorey’s) was established based on around 5500 plots of the French National Forest Inventory (NFI), distributed across the entire area of interest. In this study, we evaluate models with respect to their probabilities of correctly predicting measurements or estimations obtained from NFI plots, highlighting the spatial variability in respective model probabilities across the study area. We observed significant variability in these probabilities depending on the forest height definition used, implying that the different models inadvertently predict different types of canopy height. We also present the respective inter-model and intra-model variance estimations, enabling us to grasp where the employed models have comparable contributions but contrasting predictions. We show that topography has an important impact on the models spread. Moreover, we observed that the forest stand vertical structure, the dominant tree species, and the type of forest ownership systematically emerge as statistically significant factors influencing the model divergences. Finally, we observed that the fitted higher-order mixtures, which enabled the presented analyses, do not necessarily reduce bias or prevent the saturation of the predicted heights observed in the individual models.

Many groups now publish their own satellite maps of forest height, and these maps often disagree. With Nikola Besic and colleagues from IGN, France’s national forest inventory, we showed that rather than picking one map you can combine several with a statistical method (Bayesian model averaging) to get a better estimate and, crucially, a measure of how uncertain it is.

2024

Combining satellite images with national forest inventory measurements for monitoring post-disturbance forest height growth

Agnès Pellissier-Tanon, Philippe Ciais, Martin Schwartz, Ibrahim Fayad, Yidi Xu, François Ritter, Aurélien Truchis, and Jean-Michel Leban

Frontiers in Remote Sensing, Aug 2024

abstract In brief paper

Introduction: The knowledge about forest growth, influenced by factors such as tree species, tree age, and environmental conditions, is a key for future forest preservation. Height and age data can be combined to describe forest growth and used to infer known environmental effects. Methods: In this study, we built 14 height growth curves for stands composed of monospecific or mixed species using ground measurements and satellite data. We built a random forest height model from tree species, age, area of disturbance, and 125 environmental parameters (climate, altitude, soil composition, geology, stand ownership, and proximity to road and urban areas). Using feature elimination and SHapley Additive exPlanations (SHAP) analysis, we identified six key features explaining the forest growth and investigated how they affect the height. Results: The agreement between satellite and ground data justifies their simultaneous exploitation. Age and tree species are the main predictors of tree height (49% and 10%, respectively). The disturbed patch area, revealing the regeneration method, impacts post-disturbance growth at 19%. The soil pH, altitude, and climatic water budget in summer impact tree height differently depending on the age and tree species. Discussion: Methods integrating satellite and field data show promise for analyzing future forest evolution.

How fast a forest grows depends on the trees’ species and age and on their environment, but untangling these effects is hard. With Agnès Pellissier-Tanon and colleagues from LSCE, we combined field measurements with satellite data to build height-growth curves for French forest stands, and used machine learning to find which factors matter most. Age and species dominate, while the way a forest regenerates after disturbance, and local soil and climate, also leave a clear mark.
Estimating Canopy Height at Scale

Jan Pauls, Max Zimmer, Una M. Kelly, Martin Schwartz, Sassan Saatchi, Philippe Ciais, Sebastian Pokutta, Martin Brandt, and Fabian Gieseke

Forty-first International Conference on Machine Learning (ICML), Jun 2024

abstract In brief paper

We propose a framework for global-scale canopy height estimation based on satellite data. Our model leverages advanced data preprocessing techniques, resorts to a novel loss function designed to counter geolocation inaccuracies inherent in the ground-truth height measurements, and employs data from the Shuttle Radar Topography Mission to effectively filter out erroneous labels in mountainous regions, enhancing the reliability of our predictions in those areas. A comparison between predictions and ground-truth labels yields an MAE/RMSE of 2.43 / 4.73 (meters) overall and 4.45 / 6.72 (meters) for trees taller than five meters, which depicts a substantial improvement compared to existing global-scale products. The resulting height map as well as the underlying framework will facilitate and enhance ecological analyses at a global scale, including, but not limited to, large-scale forest and biomass monitoring.

With Jan Pauls and colleagues from the University of Münster, we built an AI model that estimates the height of trees anywhere on Earth from satellite images. Our main ideas are a training method that copes with the small location errors in the reference measurements, and the use of terrain data to avoid mistakes in mountainous areas. The result is a global tree-height map that is noticeably more accurate than previous ones.
High-resolution canopy height map in the Landes forest (France) based on GEDI, Sentinel-1, and Sentinel-2 data with a deep learning approach

Martin Schwartz, Philippe Ciais, Catherine Ottlé, Aurelien De Truchis, Cedric Vega, Ibrahim Fayad, Martin Brandt, Rasmus Fensholt, Nicolas Baghdadi, François Morneau, David Morin, Dominique Guyon, Sylvia Dayau, and Jean-Pierre Wigneron

International Journal of Applied Earth Observation and Geoinformation, Apr 2024

abstract In brief paper

In intensively managed forests in Europe, where forests are divided into stands of small size and may show heterogeneity within stands, a high spatial resolution (10–––20 m) is needed to capture the differences in canopy height. In this work, we developed a deep learning model based on multi-sensor remote sensing measurements to create a high-resolution canopy height map over the “Landes de Gascogne” forest in France, a large maritime pine plantation of 13,000 km2 with flat terrain and intensive management. This area is characterized by even-aged and mono-specific stands, of a typical length of a few hundred meters, harvested every 35 to 50 years. Our deep learning U-Net model uses multi-band images from Sentinel-1 and Sentinel-2 with composite time averages as input to predict tree height derived from GEDI waveforms. The evaluation is performed with external validation data from forest inventory plots and a stereo 3D reconstruction model based on Skysat imagery available at specific locations. We trained seven different U-Net models based on combinations of Sentinel-1 and Sentinel-2 bands to evaluate the importance of each sensor in the dominant height retrieval. The model outputs allow us to generate a 10 m resolution canopy height map of the whole “Landes de Gascogne” forest area for 2020 with a mean absolute error of 2.02 m on the test dataset. The best predictions were obtained using all available bands from Sentinel-1 and Sentinel-2 but using only one satellite source also provided good predictions. For all validation datasets in coniferous forests, our model showed better metrics than previous canopy height models available in the same region.

European forests are divided into small stands, so you need very detailed maps to tell them apart. Focusing on the Landes de Gascogne, France’s huge maritime pine plantation, my colleagues and I trained a deep-learning model on Sentinel-1 and Sentinel-2 images to map tree height at 10 metre resolution. This was an early proof that combining these satellites with AI can measure forest height accurately over a whole region.
Hy-TeC: a hybrid vision transformer model for high-resolution and large-scale mapping of canopy height

Ibrahim Fayad, Philippe Ciais, Martin Schwartz, Jean-Pierre Wigneron, Nicolas Baghdadi, Aurélien De Truchis, Alexandre d’Aspremont, Frederic Frappart, Sassan Saatchi, Ewan Sean, Agnes Pellissier-Tanon, and Hassan Bazzi

Remote Sensing of Environment, Mar 2024

abstract In brief paper

Accurate and timely monitoring of forest canopy height is critical for assessing forest dynamics, biodiversity, carbon sequestration as well as forest degradation and deforestation. Recent advances in deep learning techniques, coupled with the vast amount of spaceborne remote sensing data offer an unprecedented opportunity to map canopy height at high spatial and temporal resolutions. Current techniques for wall-to-wall canopy height mapping correlate remotely sensed information from optical and radar sensors in the 2D space to the vertical structure of trees using lidar’s 3D measurement abilities serving as height proxies. While studies making use of deep learning algorithms have shown promising performances for the accurate mapping of canopy height, they have limitations due to the type of architectures and loss functions employed. Moreover, mapping canopy height over tropical forests remains poorly studied, and the accurate height estimation of tall canopies is a challenge due to signal saturation from optical and radar sensors, persistent cloud cover, and sometimes limited penetration capabilities of lidar instruments. In this study, we map heights at 10 m resolution across the diverse landscape of Ghana with a new vision transformer (ViT) model, dubbed Hy-TeC, optimized concurrently with a classification (discrete) and a regression (continuous) loss function. This model achieves significantly higher accuracy than previously employed convolutional-based approaches (ConvNets) optimized with only a continuous loss function. Hy-TeC results show that our proposed discrete/continuous loss formulation significantly increases the sensitivity for very tall trees (i.e., \textgreater 35 m). Overall, Hy-TeC has significantly reduced bias (0.8 m) and higher accuracy (RMSE = 6.6 m) over tropical forests for which other approaches show poorer performance and oftentimes a saturation effect. The height maps generated by Hy-TeC also have better ground sampling distance and better sensitivity to sparse vegetation. Over these areas, Hy-TeC showed an RMSE of 3.1 m in comparison to a reference dataset while the baseline ConvNet model had an RMSE of 4.3 m. Hy-TeC, which was used to generate a height map of Ghana using free and open access remotely sensed data with Sentinel-2 and Sentinel-1 images as predictors and GEDI height measurements as calibration data, has the potential to be used globally.

Knowing the height of trees across whole countries is key for tracking forest carbon, degradation and deforestation. With Ibrahim Fayad and colleagues from LSCE, we developed Hy-TeC, a deep-learning model that turns optical and radar satellite images into high-resolution, wall-to-wall maps of canopy height, trained against LiDAR height measurements.

2023

Mapping forest height and biomass at high resolution in France with satellite remote sensing and deep learning

Martin Schwartz

PhD dissertation, Dec 2023

abstract In brief paper

The major role played by forests in the carbon cycle is a key component in climate change mitigation. Increasing forest biomass stocks is at the center of France’s national low-carbon strategy, which aims to achieve carbon neutrality by 2050 and is a key element of the European strategy to cut emissions by 55%. French forests are mainly divided into small parcels of a few hectares, which results in a very heterogeneous biomass distribution at the local scale. Therefore, ensuring real-time monitoring at a scale that is relevant to forest management operations requires the support of high spatial and temporal resolution data. The aim of this thesis is to use new optical (Sentinel-2), RADAR (Sentinel-1), and LiDAR (GEDI, onboard the International Space Station) satellite data, coupled with recent advances in artificial intelligence, to improve our knowledge of the dynamics of aerial biomass carbon stocks on a local scale in France.The first part of this work is devoted to the development and evaluation of a methodology focusing on the Landes de Gascogne forest, a one million-hectare maritime pine plantation in southwestern France. We used a deep learning model originally designed for medical image segmentation (U-Net) by training it to predict heights from the GEDI LiDAR sensor using optical (Sentinel-2) and RADAR (Sentinel-1) satellite imagery. Comparison with independent height measurements shows that the map obtained for the year 2020 is more accurate than all other existing products for the Landes forest, paving the way for a better understanding of small-scale forest structures.The aim of the second chapter is to create high-resolution biomass maps at a national scale using the same method as previously. The height map for France shows very good validation results, surpassing those of previous studies. We also developed allometric equations to convert the height map into biomass and volume maps at 30 m resolution. These products, which are available online under the acronyms FORMS-H, FORMS-V, and FORMS-B (FORest Multiple Satellite Height, Volume, Biomass), improve our knowledge of French forests at the local level and can inform decision processes in forest management.The third chapter explores a use case for the maps generated above. By combining the FORMS-H map with high-resolution fire polygons and local measurements of forest biomass, we obtained precise estimates of the areas burnt and the biomass affected by forest fires over the last three years. Our results reveal a significant loss of biomass in temperate forests in 2022, underlining the devastating impact of fires in these regions.The final chapter aims to obtain an annual forest growth indicator by modifying our methods. Here, we used raw satellite images, rather than time averages, to train the model to predict heights independently of the acquisition conditions. We then obtained a time series of heights in France over six years that contained an annual growth signal at plot level, particularly for fast-growing species such as the Maritime Pine. This chapter opens up an interesting avenue for the annual monitoring of the carbon budget of French forests at the local level.Through its significant contribution to the mapping of height and biomass of French forests, this thesis provides crucial information that can be used in the context of forest management in line with current climate issues.

This is my PhD thesis, which ties together much of the work here. In it I developed methods that combine radar, optical and LiDAR satellite data with deep learning to map the height and biomass of French forests at high resolution, and applied them to questions such as forest growth and fire damage. It lays the groundwork for monitoring French forest carbon from space.
FORMS: Forest Multiple Source height, wood volume, and biomass maps in France at 10 to 30 m resolution based on Sentinel-1, Sentinel-2, and Global Ecosystem Dynamics Investigation (GEDI) data with a deep learning approach

Martin Schwartz, Philippe Ciais, Aurélien De Truchis, Jérôme Chave, Catherine Ottlé, Cedric Vega, Jean-Pierre Wigneron, Manuel Nicolas, Sami Jouaber, Siyu Liu, Martin Brandt, and Ibrahim Fayad

Earth System Science Data, Nov 2023

abstract In brief paper

The contribution of forests to carbon storage and biodiversity conservation highlights the need for accurate forest height and biomass mapping and monitoring. In France, forests are managed mainly by private owners and divided into small stands, requiring 10 to 50 m spatial resolution data to be correctly separated. Further, 35 % of the French forest territory is covered by mountains and Mediterranean forests which are managed very extensively. In this work, we used a deep-learning model based on multi-stream remote-sensing measurements (NASA’s Global Ecosystem Dynamics Investigation (GEDI) lidar mission and ESA’s Copernicus Sentinel-1 and Sentinel-2 satellites) to create a 10 m resolution canopy height map of France for 2020 (FORMS-H). In a second step, with allometric equations fitted to the French National Forest Inventory (NFI) plot data, we created a 30 m resolution above-ground biomass density (AGBD) map (Mg ha−1) of France (FORMS-B). Extensive validation was conducted. First, independent datasets from airborne laser scanning (ALS) and NFI data from thousands of plots reveal a mean absolute error (MAE) of 2.94 m for FORMS-H, which outperforms existing canopy height models. Second, FORMS-B was validated using two independent forest inventory datasets from the Renecofor permanent forest plot network and from the GLORIE forest inventory with MAE of 59.6 and 19.6 Mg ha−1, respectively, providing greater performance than other AGBD products sampled over France. Finally, we compared FORMS-V (for volume) with wood volume estimations at the ecological region scale and obtained an R2 of 0.63 with an MAE of 30 m3 ha−1. These results highlight the importance of coupling remote-sensing technologies with recent advances in computer science to bring material insights to climate-efficient forest management policies. Additionally, our approach is based on open-access data having global coverage and a high spatial and temporal resolution, making the maps reproducible and easily scalable. FORMS products can be accessed from https://doi.org/10.5281/zenodo.7840108 (Schwartz et al., 2023).

Forests store carbon, but France’s are split into many small, privately owned parcels that are hard to monitor. My colleagues and I used a deep-learning model fed with Sentinel-1, Sentinel-2 and GEDI satellite data to map the height, wood volume and biomass of French forests at 10 to 30 metre resolution. These maps, which we call FORMS, give a detailed nationwide picture to support forest management and carbon accounting.
High-resolution data reveal a surge of biomass loss from temperate and Atlantic pine forests, contextualizing the 2022 fire season distinctiveness in France

Lilian Vallet^★, Martin Schwartz^★, Philippe Ciais, Dave Wees, Aurelien Truchis, and Florent Mouillot ^★ equal contribution

Biogeosciences, Sep 2023

abstract In brief paper

The frequency and intensity of summer droughts and heat waves in Western Europe have been increasing, raising concerns about the emergence of fire hazard in less fire-prone areas. This exposure of old-growth forests hosting unadapted tree species may cause disproportionately large biomass losses compared to those observed in frequently burned Mediterranean ecosystems. Therefore, analyzing fire seasons from the perspective of exposed burned areas alone is insufficient; we must also consider impacts on biomass loss. In this study, we focus on the exceptional 2022 summer fire season in France and use very high-resolution (10 m) satellite data to calculate the burned area, tree height at the national level, and subsequent ecological impact based on biomass loss during fires. Our high-resolution semi-automated detection estimated 42 520 ha of burned area, compared to the 66 393 ha estimated by the European automated remote sensing detection system (EFFIS), including 48 330 ha actually occurring in forests. We show that Mediterranean forests had a lower biomass loss than in previous years, whereas there was a drastic increase in burned area and biomass loss over the Atlantic pine forests and temperate forests. High biomass losses in the Atlantic pine forests were driven by the large burned area (28 600 ha in 2022 vs. 494 ha yr−1 in 2006–2021 period) but mitigated by a low exposed tree biomass mostly located on intensive management areas. Conversely, biomass loss in temperate forests was abnormally high due to both a 15-fold increase in burned area compared to previous years (3300 ha in 2022 vs. 216 ha in the 2006–2021 period) and a high tree biomass of the forests which burned. Overall, the biomass loss (i.e., wood biomass dry weight) was 0.25 Mt in Mediterranean forests and shrublands, 1.74 Mt in the Atlantic pine forest, and 0.57 Mt in temperate forests, amounting to a total loss of 2.553 Mt, equivalent to a 17 % increase of the average natural mortality of all French forests, as reported by the national inventory. A comparison of biomass loss between our estimates and global biomass/burned areas data indicates that higher resolution improves the identification of small fire patches, reduces the commission errors with a more accurate delineation of the perimeter of each fire, and increases the biomass affected. This study paves the way for the development of low-latency, high-accuracy assessment of biomass losses and fire patch contours to deliver a more informative impact-based characterization of each fire year.

France’s 2022 summer brought an exceptional fire season, with flames reaching temperate and Atlantic pine forests that rarely burn. Using high-resolution (10 m) satellite images, Lilian Vallet and I mapped where forests burned and how much woody biomass, and therefore carbon, was lost. We found the damage was far larger than usual outside the Mediterranean, a sign of how climate change is pushing fire into new regions.