Forestry Engineering student building production-grade geospatial ML pipelines.
UAV photogrammetry · Sentinel-1/2 remote sensing · deep learning for forest inventory
Córdoba, Spain

The bottleneck in forestry isn't data — it's tooling. UAV imagery, Sentinel time series, LiDAR and NDVI are already free. What's missing are pipelines a forester can actually run, on real datasets, on hardware they own. That's the gap I work in.

I'm a forestry engineering student with a strong interest in geospatial analysis and machine learning. Forest management both generates and depends on very large volumes of data — UAV surveys, satellite time series, LiDAR, field inventories — and a good part of it stays underused for lack of accessible tooling. That is where I focus my work: building pipelines that bring technology into field data collection and make it faster, automate the processing and analysis of that data, and return results that are easy to interpret for specialists and non-specialists alike.

Most of it happens in the dehesas of southern Spain: flying the UAV campaigns, walking the plots, and then sitting down with the imagery. I enjoy working at that intersection between forestry and code, and I'm always glad to talk with people working on similar problems.

Every figure below is the one I would defend in front of someone who wanted to check it. Where a model loses to a naive baseline, the repository says so and leads with it. Where an uncertainty cannot be estimated honestly, no number is invented to fill the gap. Each project documents its own limitations, because a result you cannot see the edges of is not a result you can use.

Instance segmentation of individual tree crowns on UAV orthomosaics. Run over 14,506 ha at 25 cm/px, producing an inventory of 357,185 crowns with per-tree area, diameter and centroid, exported to GeoPackage for QGIS. Model accuracy is mask mAP@50-95 = 42.3 % on the validation split: good enough for stand-level structure, not for an exact individual count — the inventory is a lower bound, and the repository says so. Out-of-core tiling with one model instance per worker thread and explicit accounting of failed tiles, so a partial run can never be mistaken for a complete one.

Sentinel-2 L2A pipeline on the Microsoft Planetary Computer STAC catalogue — no credentials, no subscriptions, COG range reads instead of full-scene downloads. AOI-driven scene search, cloud filtering, scene de-duplication across processing baselines, automatic merging of AOIs that span several MGRS tiles, and six spectral indices (NDVI, SAVI, EVI, NBR, NDRE, NDWI) with a per-scene validation dashboard.

Sentinel-1 + Sentinel-2 forest-disturbance monitoring: STAC-native Dask cube, omnibus Wishart change detection, Theil-Sen/Mann-Kendall trends. Sierra de la Culebra 2022 fire → 4,829 ha of error-adjusted forest loss, scored against Hansen Global Forest Change at precision 0.75 / recall 0.99 / F1 0.85. The raw map says 6,380 ha; the Olofsson (2014) adjustment corrects it down, which is the point of adjusted reporting. No confidence interval is quoted, because a wall-to-wall reference cannot support one.

Local, privacy-first image annotation platform (FastAPI + React + SAM 2.1). Three-level embedding cache — in-memory, on-disk feature tensors, then recompute — runs on 4 GB VRAM; exports to YOLO-seg, YOLO-det and COCO. No telemetry, no outbound requests, no image ever leaves the machine. This is the tool that produced the training data for the crown segmentation model above.

Weekly olive-oil origin price forecasting: bimodal LSTM over climate and macroeconomic series, FastAPI service, Docker. Reported as a negative result. Against a persistence baseline (MAE 0.057 €/kg) the model scores 0.073 — it does not beat repeating last week's price, and the README leads with that. Building the baseline is also what surfaced a two-week feature-alignment bug that had been inflating the error.

Internships in geospatial ML / precision forestry / remote sensing / computer vision (Spain or remote) · R&D collaborations on operational forest monitoring · Conversations at the intersection of forestry, earth observation & applied AI