Marie Skłodowska-Curie Postdoctoral Fellow
Image and Signal Processing (ISP) group
University of Valencia
digo [at] uv.es
Reviewer: CVPR’24’25’26, ECCV’26, CVPR Workshop EarthVision’24’25’26
Outstanding Reviewer: CVPR’26, ECCV’26
My work sits between computer vision and environmental science: I design models that turn satellite imagery into ecological measurements, such as individual tree locations and species, ecosystem types and degradation patterns. I am driven by applications with real environmental impact, and by the technical challenges they raise: learning from noisy labels, low-data regimes and time-series analysis.
ECCV, 2026
We frame tree counting as spatial density matching with unbalanced optimal transport. This handles both isolated trees and dense canopies, and a self-correction step uses transport residuals to clean up noisy labels during training.
ECCV, 2026
Canopy change is usually treated as binary change detection. We release the Canopy Height Change dataset, with 3 m continuous height differences and per-pixel uncertainty over 10,598 km² of Spain, and introduce uncertainty-aware change regression with matching metrics.
ECCV, 2026
Labels from sources like OpenStreetMap rarely line up with the imagery. A spatial transformer learns to realign the labels while the segmentation network trains, with a self-supervised loss that stops the network from learning the misalignment.
Nature, 2026
A 10 m map of six semi-natural wetland types and land-use disturbance across 38 European countries. It shows that wetlands are highly fragmented, that about 20% are affected by human activity, and where restoration under the EU Nature Restoration Law matters most.
Journal of Remote Sensing, 2026
A global model that detects large individual trees in 3 m PlanetScope imagery. Crowns are modeled as scalable Gaussian kernels, giving both tree centers and tree cover maps. Training uses billions of points extracted automatically from airborne LiDAR, so trees are found both inside and outside forests.
Journal of Remote Sensing, 2026
Counting trees nationally across Rwanda and Tanzania from freely available PlanetScope, Sentinel-1 and Sentinel-2 data. We estimate about 5.2 billion trees, roughly half of them outside areas mapped as "tree cover".
Nature Sustainability, 2024
We map 0.6 billion individual farmland trees across India and track them over a decade. About 11% of large trees mapped in 2010/2011 were gone by 2018, and more than 5 million more disappeared between 2018 and 2022.
ECCV, 2024
project page / arXiv / paper
DRIFT is a dataset of aerial and satellite imagery from five countries with three forest regression tasks, built to study domain shift. We propose manifold diffusion for regression, a strong baseline for adapting at test time with few labels.
ECCV Workshops, 2024
Species labels for individual trees are automatically mined from public forest inventories by matching field plots with trees detected in aerial imagery, with almost no human work. Adding noisy and even unlabeled samples helps, which points toward species mapping at scale.
Communications Earth & Environment, 2024
Carbon stocks of every tree in Rwanda in 2008 and 2019. Smallholder farmers planted about 50 million new trees, and agroforestry and forest restoration together could offset a substantial share of national emissions.
CVPR Workshops (EarthVision), 2022
Best Student Paper Award
A retrieval-inspired framework that trains on several land-use datasets at once and classifies scenes from unseen datasets and unseen classes with only a few reference images, going beyond standard few-shot and cross-domain settings.