Caleb Robinson

Excited to announce that TerraKit and TerraTorch from @IBMResearch have joined the TorchGeo organization - github.com/torchgeo!! Together with TorchGeo, these python packages provide geospatial ML datasets, foundation models, fine-tuning tools, and more!

@jo_chemla @GoogleDeepMind @WebGPU @STACspec Thanks @jo_chemla, I think we're coauthors by the way :) -- arxiv.org/abs/2602.19608

@GoogleDeepMind @WebGPU First time I see an interactive change-detection app based on pixel vector embeddings (rather than patch sim etc) + running fully in-browser congrats! Would love to be able to user-define the S2 pair (BYOD COG 🤞, or from @STACspec fetch)

Amazing deep-dive! Great findings, porting embeddings compression techniques from patch to pixel. Based on AlphaEarth Foundations from @GoogleDeepMind In-browser training/inference on Deltabit using @WebGPU calebrob.com/deltabit/ PCA8 vec Diff + nice idea mapping PCA3 to RGB

@nichm @isaaccorley_ Cool!! Try a _very_ few number of change and no change samples, then iterate (add more samples and retrain) based on where you see the model making mistakes

@calebrob6 @isaaccorley_ pretty cool. i was lazy and only changed 54 samples but it still mostly got this golf course.

New post with @isaaccorley_ : DeltaBit -- dense, satellite imagery-based change detection with AlphaEarth embeddings, running entirely in the browser. Check it out at geospatialml.com/posts/change-d…

This lets users label a few examples of change → train a linear model → see predictions interactively as they scroll around Seattle → download change polygons. No backend services or desktop software needed! Try it at calebrob.com/deltabit/

This is pt. 3 of our "Compressing Earth Embeddings" series. We: - Compress per-pixel *differences* between two years of embeddings and store the top 8 PCA components as XYZ GeoTIFF tiles - Use GeoTIFF.js to decode in-browser - Use WebGPU compute shaders to train/inference

@MapMakinMeyers I'm so sorry Justin :(

She's making her last trip home. 24 hours to live. Where would you go. What would you do. We are devastated.

Working on transfer learning and foundation models in remote sensing? We invite you to submit abstracts to our session IN005 at #AGU25 (agu.confex.com/agu/agu25/prel……) organized by @sherwang (MIT),@hannah_kerner (Arizona State), Siddharth Sachdeva, and me (Stanford).

Researchers at @Microsoft’s AI for Good Lab, in collaboration with @WildNatureInst, have developed GIRAFFE, an open-source AI-driven system to automate the identification of individual giraffes. Learn more about our new AI-powered tool. msft.it/6017SQgyW

I'm hiring for a machine learning data scientist & research assistant for summer 2025! Join me on a project on invasive species management with an innovative startup doing on-the-ground removal of environmentally destructive invasives. Paid, full-time w/ possibility to extend.

An expanded version of our TorchGeo paper has been accepted to ACM TSAS @TheOfficialACM In addition to even more benchmarks on remote sensing datasets, we provide a thorough survey of open-source libraries for deep learning for geospatial data @calebrob6 @BDataScientist

@Ritwik_G Here's a small experiment using MOSAIKS features and KNN-3 (notebook gist.github.com/calebrob6/7495…). Looks like there is some interesting space for showing gains vs. simple methods with few training samples. Would love to see 90+ acc with a handful of examples!

@calebrob6 There have been papers that have evaluated on EuroSAT in few-shot settings, but there’s no standardized dataset for it. Let’s talk about it!

Remote sensing people, we need to stop using EuroSAT as a benchmark. It is saturated and provides no meaningful differentiation between new methods.

NOAA has posted satellite imagery from the areas affected by the Eaton, Hurst, and Palisades wildfires. The imagery is available through Maxar's Open Data program to support the emergency response community and the public: #9.98/34.1904/-118.4094" target="_blank" rel="nofollow noopener">storms.ngs.noaa.gov/storms/2025_er… Disaster-specific @FEMA resources are available here: fema.gov/disaster/4856

@mouthofmorrison @dmarvs @earth_chris And 6,3,1 gets you pretty close :) -- we reproduced here aka.ms/los_angeles_ea…

@mouthofmorrison @dmarvs @earth_chris The raw data is in their open data program if you're curious! -- #events/WildFires-LosAngeles-Jan-2025/or2a/016511482010_01/016511482010_01_P002_SWR/" target="_blank" rel="nofollow noopener">qoomon.github.io/aws-s3-bucket-…

Pro tip: never share satellite imagery without attribution in general but REALLY try to avoid sharing Maxar money shot SWIR data without attribution

@yacineMTB @growing_daniel Yes, you can see before/after satellite imagery here aka.ms/los_angeles_pa…. There are ~2k destroyed buildings visible in this scene

@growing_daniel Wait is that a real picture on the left

Palisades vs. Hiroshima

The Climate Change AI workshop at NeurIPS tomorrow will be livestreamed for free, even if you weren't able to register for NeurIPS (which was oversubscribed this year)!

🚨Our new review paper "Better, not just more: Data-centric ML for Earth Observation" 🌐🤖🛰️🔍 has finally been published in GRSM. Shifting the focus from models to data offers new possibilities to improve accuracy, generalization, and real-world impact: ieeexplore.ieee.org/document/10740… Joint work with @MarcCoru, Caroline Gevaert, @MKampffmeyer, @jeferssonalex, @MariaVakalopou1, @RonnyHaensch, Stine Hansen, @Keiller, Jonathan Prexl, @devistuia Arxiv version: arxiv.org/abs/2312.05327

@MaxLenormand If you don't need very recent -- there is 60cm NAIP imagery from 2022 on the Planetary Computer (RGBNIR)

Distraction about the stuff going on today to all my American friends: Anyone know where I can get some high res optical imagery of New York? Quick search didn't find anything