Land-Cover Classification from Sentinel-2

Why build this

My internship work involves satellite imagery I can't share publicly. This project uses the same core skills (multispectral imagery, transfer learning, careful evaluation, deployment) on open data, so anyone can poke at the result.

The goal wasn't the highest possible accuracy on a well-studied benchmark. It was to take a model all the way to a place where a stranger can use it, in a browser tab with no server, and to be honest about what it can and can't do.

Live demo

What's in this satellite tile?

A ResNet-18 I fine-tuned on Sentinel-2 imagery classifies land cover into 10 classes and shows where it's looking.

model

ResNet-18 · int8

size

~11 MB

runtime

ONNX Runtime Web

Model in training

This demo goes live once the model finishes training and evaluation. Check back soon.

Data

EuroSAT is 27,000 labeled 64×64 RGB tiles from ESA's Sentinel-2 satellite, covering 10 land-cover classes across 34 European countries. I use a stratified 80/10/10 split with a fixed seed. The validation set picks the checkpoint, and the test set is touched exactly once.

Model

• Backbone: ResNet-18 initialized from ImageNet, the whole network fine-tuned.
• Input: tiles upsampled to 128×128, with the stem max-pool removed so the final feature map is 8×8 instead of 4×4. That's the resolution of the attention heatmap, so it matters for the demo.
• Augmentation: random 90° rotations and flips (satellite tiles have no canonical "up") plus mild color jitter.
• Training: AdamW, one-cycle LR schedule, label smoothing 0.1, bf16 autocast.

Shipping it to the browser

1. ONNX export with two outputs. Besides logits, the graph computes a class activation map (Zhou et al., 2016). The final conv features are weighted by the classifier weights for each class. The heatmap costs one extra einsum and no second pass.
2. Static int8 quantization (QDQ, per-channel weights) calibrated on 300 training tiles. This makes the model ~4× smaller with a negligible accuracy change (see below).
3. Web Worker + ONNX Runtime Web (WASM). Inference runs off the main thread, so the page never janks. The model downloads with a progress bar, then the browser caches it.
4. Vegetation view. Real NDVI needs the near-infrared band, which RGB tiles don't have. The demo computes VARI, an RGB-only proxy, and labels it as such.

Results

What I'd do next

• Train on all 13 spectral bands (EuroSAT-MS) with a modified first conv, to compare RGB-only against multispectral.
• Rerun with a spatially blocked split to measure how much the random split overstates accuracy.
• Try the WebGPU execution provider for fp16 inference on capable devices.