🛰️ AstroRoot

Train your own model

The neural network can't be trained inside the browser — that needs a GPU. So AstroRoot splits the job: you label examples here, train in the cloud (a free Colab or CyVerse notebook), then load the finished .onnx back into AstroRoot.

1. Label roots. Load a photo, then trace each root by clicking along it (click to add points, double-click to finish a root). Each trace is saved as an RSML polyline.
   Undo point Finish root Clear image 0 roots traced
2. Build the training set. Add each labelled image to your dataset, then export it as a zip of image + RSML pairs (the exact format RootNav 2.0 trains on).
   Add this image to dataset Export training set (.zip) 0 images in dataset
3. Train in the cloud. Open the training notebook, upload your zip, run it (≈30–60 min on a free GPU). It fine-tunes the Arabidopsis model and gives you a model.onnx.
   Open training guide
4. Use it. Back on the Single/Batch tabs, choose My own model (.onnx)… and pick your file. It now traces roots in your style.

About AstroRoot

A free, install-free root-image analyser for students and teachers. It brings together the friendly, correct-by-hand feel of RootNav 1 with the automatic tracing of RootNav 2.0's deep-learning model — running entirely in your browser, so student photos never leave the device.

• Auto engine: RootNav 2.0 stacked-hourglass model, exported to ONNX, run with ONNX Runtime Web. (Until a model is loaded, a transparent classical baseline is used.)
• Interactive layer: RootNav 1-style hand correction of the traced result.
• Outputs: length, tips, branch points, gravitropic angle; CSV + RSML export.
• Calibration: the AstroCalibration marker sets pixels → millimetres.

New here? The hands-on tutorial walks through every feature. Start with the Synthetic — clean, known answer demo roots in the picker above: a simple shape drawn at a known scale, so the measurements can be checked against an exact expected value (unlike the real, deliberately diverse NASA/ABRS plates).

Built on RootNav 2.0 and RootNav 1 (both BSD-3-Clause, University of Nottingham). Part of the AIRI astrobotany program.