Periodic Reporting for period 2 - TREES4CLIMA (Enabling carbon accounting of trees on farms for agroforestry-based climate action)
Période du rapport: 2023-09-01 au 2024-08-31
Agroforestry systems integrate trees, crops and livestock on farms, providing adaptation and mitigation to climate change and livelihood opportunities. Despite its benefits, agroforestry is often not accounted for in measurement, reporting and verification (MRV) systems, which are a requirement for fulfilling countries’ climate goals under the Paris Agreement. Furthermore, new regulations such as EUDR may have collateral damage in agroforestry systems due to its resemblance to forests. If agroforestry is not included, incentives for leveraging on-farm trees for addressing climate change will be limited.
TREES4CLIMA addressed three shortcomings in scientific research that prevent the adoption of agroforestry: 1) Lack of accessible approaches for the representation of agroforestry in MRV systems; 2) Shortfall of data and tailored methods for carbon accounting in agroforestry systems; 3) Institutional barriers that withhold agroforestry from access to MRV and climate finance. The overall research objective was to develop and test robust, cost-effective approaches to account for carbon in agroforestry systems and enable innovative environments for climate finance, and the creation of practical pathways to integrate agroforestry into MRV and regulatory frameworks that unlock incentives.
The project has achieved its objectives and milestones for the period, including A) the classification and representation of agroforestry systems by accessible remote sensing and deep learning; B) advancing methods for rapid quantification of carbon in biomass on farms using novel computer vision and field images; C) quantification of soil carbon across land uses; and D) applying innovative solutions to link individual trees and climate mechanisms. TREES4CLIMA demonstrates that credible, scalable representation for agroforestry is feasible with accessible tools, reduces the cost of carbon accounting, and clarifies pathways for climate finance.
TREES4CLIMA addressed three shortcomings in scientific research that prevent the adoption of agroforestry: 1) Lack of accessible approaches for the representation of agroforestry in MRV systems; 2) Shortfall of data and tailored methods for carbon accounting in agroforestry systems; 3) Institutional barriers that withhold agroforestry from access to MRV and climate finance. The overall research objective was to develop and test robust, cost-effective approaches to account for carbon in agroforestry systems and enable innovative environments for climate finance, and the creation of practical pathways to integrate agroforestry into MRV and regulatory frameworks that unlock incentives.
The project has achieved its objectives and milestones for the period, including A) the classification and representation of agroforestry systems by accessible remote sensing and deep learning; B) advancing methods for rapid quantification of carbon in biomass on farms using novel computer vision and field images; C) quantification of soil carbon across land uses; and D) applying innovative solutions to link individual trees and climate mechanisms. TREES4CLIMA demonstrates that credible, scalable representation for agroforestry is feasible with accessible tools, reduces the cost of carbon accounting, and clarifies pathways for climate finance.
Below there is a summary of the tasks performed across the four objectives of TREES4CLIMA:
A. Classify and represent agroforestry systems with remote sensing and deep learning
Downloading and processing satellite data. I downloaded high-resolution PlanetScope imagery via the NICFI programme. A standardized preprocessing workflow—mosaicking and normalization— was applied based on established protocols to produce 3 m resolution PlanetScope mosaics.
Obtaining training and validation data. During field campaigns, drone images and orthophotos were acquired across hedgerows, silvopastures, multistrata, shaded systems, and fallows. Farm geolocations were collected, as well as polygons under the Peruvian platform for EUDR verification
Designing a deep learning framework. I developed a dual-branch model, agroforestry_detection, which ingests satellite imagery and forest-loss layers in parallel, learns high-level semantic features, and fuses them for agroforestry classification.
B. Advance methods for rapid quantification of on-farm biomass carbon
Imaging and acquiring data. Across orchards and agroforestry sites, I collected UAV-mounted RGB imagery plus ground-level captures from smartphone cameras and smartphone LiDAR. Data were acquired in orchards in California, silvopastoral systems in Colombia, and fallows and multistrata systems in Peru.
Reconstructing 3D models. I generated agroforestry 3D models using a hybrid pipeline that combines structure-from-motion with state-of-the-art neural methods. These approaches were tested first in California orchards, then applied across hedgerows, silvopastures, multistrata systems, and shaded coffee in Latin America.
Estimating and modeling biomass. In parallel, I collected biomass and conducted forest inventories to train and evaluate machine-learning models. I derived plot and tree-level structural metrics and trained/evaluated supervised models to estimate biomass and benchmark predictions against field observations.
C. Quantification of carbon in soils
Field sampling and soil processing. Fieldwork covered soil sampling in georeferenced plots in the Maya Biosphere Reserve across protected forest, managed forest, and grassland.
Laboratory analyses. Soils were analyzed for SOC using Walkley–Black wet oxidation.
Predicting SOC. Linear regressions for SOC prediction—aimed at explaining carbon variation across land transitions.
D. Innovative solutions to link individual trees and climate finance
Obtaining high-resolution orthophotos. I downloaded 25 cm GSD orthophotos from Spain’s PNOA program and co-registered airborne LiDAR.
Leveraging National Forest Inventory (NFI) data. NFI data were obtained from the 4th Spanish National Forest Inventory. I developed matching_trees to address the misalignment between NFI-recorded tree positions and aerial-image features.
Implementing a deep learning pipeline to detect and classify individual trees. We introduced a pipeline that automatically matches field observations with predictions on aerial photography from an ensemble of pretrained models for individual-tree detection and segmentation.
Training undertaken
TREES4CLIMA allowed me to acquire training on: Project management; data management; advanced statistics and data science at UC Davis; deep learning courses at the University of Copenhagen; summer schools on 3D modeling, field and lab methods; pedagogics and research-based teaching at UC Davis and the University of Copenhagen; advancing communication skills inside and outside academia; and networking and proposal writing, among others.
Exploitation and dissemination of results
TREES4CLIMA’s results have been disseminated at AGU, IUFRO, ECCV, and ESA BioSpace25 and ESA Living Planet conferences, and within partner networks at CIFOR-ICRAF and CIAT. One scientific article has already been published in Ecological Informatics, with additional manuscripts in preparation. Community feedback sessions were held in San Martin for non-academic audiences. Results also informed research-based teaching and courses at the University of Copenhagen, and a short-scientific mission at the University College London (Department of Geography).
A. Classify and represent agroforestry systems with remote sensing and deep learning
Downloading and processing satellite data. I downloaded high-resolution PlanetScope imagery via the NICFI programme. A standardized preprocessing workflow—mosaicking and normalization— was applied based on established protocols to produce 3 m resolution PlanetScope mosaics.
Obtaining training and validation data. During field campaigns, drone images and orthophotos were acquired across hedgerows, silvopastures, multistrata, shaded systems, and fallows. Farm geolocations were collected, as well as polygons under the Peruvian platform for EUDR verification
Designing a deep learning framework. I developed a dual-branch model, agroforestry_detection, which ingests satellite imagery and forest-loss layers in parallel, learns high-level semantic features, and fuses them for agroforestry classification.
B. Advance methods for rapid quantification of on-farm biomass carbon
Imaging and acquiring data. Across orchards and agroforestry sites, I collected UAV-mounted RGB imagery plus ground-level captures from smartphone cameras and smartphone LiDAR. Data were acquired in orchards in California, silvopastoral systems in Colombia, and fallows and multistrata systems in Peru.
Reconstructing 3D models. I generated agroforestry 3D models using a hybrid pipeline that combines structure-from-motion with state-of-the-art neural methods. These approaches were tested first in California orchards, then applied across hedgerows, silvopastures, multistrata systems, and shaded coffee in Latin America.
Estimating and modeling biomass. In parallel, I collected biomass and conducted forest inventories to train and evaluate machine-learning models. I derived plot and tree-level structural metrics and trained/evaluated supervised models to estimate biomass and benchmark predictions against field observations.
C. Quantification of carbon in soils
Field sampling and soil processing. Fieldwork covered soil sampling in georeferenced plots in the Maya Biosphere Reserve across protected forest, managed forest, and grassland.
Laboratory analyses. Soils were analyzed for SOC using Walkley–Black wet oxidation.
Predicting SOC. Linear regressions for SOC prediction—aimed at explaining carbon variation across land transitions.
D. Innovative solutions to link individual trees and climate finance
Obtaining high-resolution orthophotos. I downloaded 25 cm GSD orthophotos from Spain’s PNOA program and co-registered airborne LiDAR.
Leveraging National Forest Inventory (NFI) data. NFI data were obtained from the 4th Spanish National Forest Inventory. I developed matching_trees to address the misalignment between NFI-recorded tree positions and aerial-image features.
Implementing a deep learning pipeline to detect and classify individual trees. We introduced a pipeline that automatically matches field observations with predictions on aerial photography from an ensemble of pretrained models for individual-tree detection and segmentation.
Training undertaken
TREES4CLIMA allowed me to acquire training on: Project management; data management; advanced statistics and data science at UC Davis; deep learning courses at the University of Copenhagen; summer schools on 3D modeling, field and lab methods; pedagogics and research-based teaching at UC Davis and the University of Copenhagen; advancing communication skills inside and outside academia; and networking and proposal writing, among others.
Exploitation and dissemination of results
TREES4CLIMA’s results have been disseminated at AGU, IUFRO, ECCV, and ESA BioSpace25 and ESA Living Planet conferences, and within partner networks at CIFOR-ICRAF and CIAT. One scientific article has already been published in Ecological Informatics, with additional manuscripts in preparation. Community feedback sessions were held in San Martin for non-academic audiences. Results also informed research-based teaching and courses at the University of Copenhagen, and a short-scientific mission at the University College London (Department of Geography).
TREES4CLIMA enhanced innovation capacity by delivering open, accessible deep-learning pipelines for agroforestry mapping, 3D reconstruction, and individual tree–level monitoring. I strengthened field research, data analysis, project management and interdisciplinary collaboration, building a tangible portfolio, and directly improved my career prospects and employability. TREES4CLIMA tools advance the state of the art by enabling low-cost, scalable carbon accounting and structural monitoring across field, farm, and landscape scales; integrating 2D/3D computer vision with field measurements; and linking individual tree monitoring directly to finance and policy. By preventing the blind spots that arise when agroforestry is unmapped—critical for compliance under regulations such as the EUDR—the work reduces risk and increases transparency. Accessible monitoring pipelines for agroforestry can unlock market opportunities by lowering transaction costs for agroforestry projects, making nature-based solutions investable and auditable, and helping producers and cooperatives access climate-finance revenue streams. At EU level, the project contributes to competitiveness and leadership in sustainable land systems. Collectively, these advances open new research avenues across sectors and scales and enrich the agroforestry discipline.