A fundamentals-first path from geospatial Python to production remote-sensing ML: the engineering rigor, SAR depth, and portfolio that hiring managers at Planet, ICEYE, Overstory, Pachama, and JPL actually screen for.
Weeks 1–3. Rebuild from first principles at speed: how EO data is stored and served in 2026, what a reflectance or backscatter number actually means, and how to get a clean Sentinel-1 image without trusting a black box.
Every employer's JD lists GDAL, rasterio, xarray, GeoPandas, STAC, and dask. The goal this week is to stop thinking of imagery as files and start thinking of it as lazily-evaluated arrays with a CRS.
A tested cube.py module plus a notebook that produces a true-colour, cloud-masked monthly composite over the AOI. README explains provider choice and the COG/Zarr decision.
dNBR is the lingua franca of post-fire assessment, and a surprising number of candidates can compute it but not explain why NIR rises and SWIR falls with healthy vegetation, or why a top-of-atmosphere dNBR is a different product.
Burn-severity notebook + a one-page results note comparing your dNBR classes against MTBS with an honest accuracy table. This becomes section 1 of Portfolio Project A.
This is the week that separates "has used Sentinel-1" from "understands SAR". Interviewers at ICEYE, Capella, and Umbra will ask about layover, speckle statistics, and why you terrain-corrected before they ask about your model.
sar_prep.py (HyP3 + OPERA fetch, dB conversion, speckle filtering) with tests, and a notebook "three RTCs, one scene" explaining differences. Short blog-style note: what C-band backscatter sees after a fire.
Weeks 4–6. Most production EO systems still run on gradient boosting and careful sampling design, not transformers. Learn to do that properly, then add forest structure from lidar and change detection from radar time series.
Spatial autocorrelation makes random train/test splits lie. A candidate who shows spatially blocked cross-validation and area-weighted accuracy with confidence intervals is instantly credible.
Forest mask pipeline with spatial CV report, area estimates with 95% CI, and an MLflow experiment log. This is Portfolio Project B's foundation.
Carbon companies (Pachama, Sylvera, Chloris, CTrees) hire on exactly this: fusing GEDI footprints with wall-to-wall imagery to map canopy height and biomass, with defensible uncertainty.
Canopy height + AGBD maps with uncertainty rasters and a validation notebook. Portfolio Project B complete: "Forest structure and carbon loss, AOI X, 2019–2025".
Cloud-free change detection is the whole reason SAR exists in forestry and fire. RADD, DIST-ALERT-S1 and most commercial deforestation alerts are built on the methods in this week.
sar_change.py with two detectors and a comparison notebook. Public write-up #1: "Detecting burned area through cloud with Sentinel-1, a baseline and where it fails".
Weeks 7–9. Segmentation networks, then geospatial foundation models (Prithvi-EO-2.0, Clay, TerraMind, AlphaEarth embeddings), then SAR-specific deep learning and a first contact with InSAR and NISAR.
U-Net on HLS burn scars is the canonical fine-tuning task, it's what NASA-IMPACT used to demonstrate Prithvi. Build the baseline yourself before touching the foundation model so you can judge whether it actually helps.
Trained U-Net with a reproducible config, metrics table, and prediction CLI. Portfolio Project A now has a learned model alongside the index baseline.
In 2026 JDs increasingly say "experience fine-tuning geospatial foundation models" and "embedding-based workflows". Knowing when a GeoFM beats a U-Net, and when it's just slower, is the skill, not the model name.
A benchmark notebook and Public write-up #2: "U-Net vs Prithvi vs Clay vs embeddings for burn-scar mapping: accuracy, labels, and cost". Portfolio Project C.
NISAR provisional products have been flowing since July 2026 and almost nobody on the job market has touched them yet. Coherence, polarimetry, and SAR-specific DL are the differentiators for SAR-vendor roles.
Coherence-augmented change detector, a SAR flood model, and a short NISAR-vs-Sentinel-1 comparison note, a timely, rare portfolio piece.
Weeks 10–12. Turn the pieces into a near-real-time wildfire monitoring service, containerize and schedule it on cloud infrastructure, then package the whole portfolio and run a targeted job search.
Overstory, Pano AI, utilities, and the Earth Fire Alliance ecosystem want people who can stitch active-fire detections, perimeters, and post-fire imagery into a pipeline that runs unattended and degrades gracefully.
A working firewatch package that runs end-to-end on a laptop for one region, with tests and a sample report.
The single biggest gap between scientists and "ML engineers" in JDs: Docker, orchestration, object storage, and CI. A week here is worth more to your hireability than another model.
Deployed, scheduled firewatch with an architecture diagram in the README and a public demo map. Public write-up #3: "From FIRMS hotspot to damage report in under an hour".
Four projects, three write-ups, and a deployed service are more than most applicants bring. This week makes them legible to a recruiter in 90 seconds and to a hiring manager in 10 minutes.
Polished portfolio, updated resume, ten applications sent, and a recurring weekly job-search block on the calendar.
Use these as a checklist at the end of each phase and as prep in week 12.
Wildfire: Overstory, Pano AI, Earth Fire Alliance / Muon Space, Vibrant Planet, Cornea, Technosylva, utilities' wildfire-risk teams (PG&E, SCE).
Forestry / carbon: Pachama, Sylvera, Chloris Geospatial, CTrees, Planet (Forest Carbon), Kayrros, Space Intelligence.
SAR: ICEYE, Capella, Umbra, EarthDaily (ex-Descartes Labs), NASA JPL (OPERA/NISAR), ASF, DLR, Privateer.
Platforms: Google (Earth Engine/AlphaEarth), Microsoft Planetary Computer Pro, Development Seed, Element 84, Floodbase.
Watch: Geospatial Jobs newsletter · Pangeo jobs · spacecrew
| Week | Phase | Portfolio thread | Public output |
|---|---|---|---|
| 01–03 | Foundations | A: burn severity (index baseline) · SAR prep | Note: C-band after fire |
| 04–06 | Core analytics | B: forest structure & carbon loss · SAR change detectors | Write-up #1: S1 burned area |
| 07–09 | Deep learning | A: learned model · C: GeoFM benchmark · NISAR note | Write-up #2: U-Net vs GeoFMs |
| 10–12 | Production | D: firewatch NRT service, deployed | Write-up #3 · portfolio · applications |