RUNS 100% IN YOUR BROWSER · OFFLINE
Supervised Classification · Live Training · Zero Install

Train a real satellite-image classifier — by painting labels and watching it learn

This is the full machine-learning loop, live and local: paint a few example pixels of each land-cover class onto a synthetic multi-band scene, train an actual softmax classifier by gradient descent in your browser, and watch it classify the entire image — then read the accuracy, the loss curve and the per-class scores like a real EO analyst. Nothing is downloaded; nothing leaves your machine.

1 Label classes — paint examples

Pick a class, then click or drag on the scene to paint training pixels. A dozen per class is enough to start — more and varied samples improve the model.
2

2 The scene & the classification

Paint a few pixels of each class, then Train.

3 Results & metrics

Train accuracy
Held-out accuracy
What am I looking at? Each pixel carries 4 synthetic "bands" (like red, green, blue, near-infrared). The model learns weights that separate your labelled classes in that 4-D colour space, then colours every pixel by its best guess. Held-out accuracy tests it on pixels it never trained on — the honest number.

1 · It really trains

A softmax (multinomial logistic) classifier over 4 bands, optimised by mini-batch gradient descent right here in JavaScript. The loss curve is real; watch it fall.

2 · Labels are the work

Just like real EO: the model is trivial, the labels are everything. Sparse or biased painting gives a confused map — the lesson every practitioner learns.

3 · Bands beyond sight

Flip to the NIR false-colour view: vegetation glows because it reflects near-infrared. The model sees all 4 bands even where your eye sees three — why multispectral matters.

4 · Honest evaluation

20% of your labels are held out from training. Train accuracy always flatters; the held-out number is the one an analyst trusts — and the gap between them is overfitting, made visible.

From concept demo to sovereign capability. This lab teaches the real loop on synthetic data, offline, on any laptop — the exact model for classified or air-gapped classrooms. The production path is the same shape at scale: swap synthetic tiles for real multispectral imagery, the in-browser trainer for a proper framework, and add domain labels curated with subject-matter experts. That progression — from this teaching lab to a domain-specific, sovereign EO-AI curriculum built on a nation's own imagery — is exactly the kind of co-development track LASS TECH proposes to build WITH a partner agency, not sell off the shelf.

Scope note: the scene is procedurally generated to teach the method; this is a learning environment, not a trained operational product.