Get in touch Menu
Get in touch
The TESSERA project open in VS Code, showing per-channel output-versus-target histograms from the test set. The models form narrow spikes where the target is spread out: the plot title calls it colour collapsing to a flat tone.
Exhibit 00 Photograph of the test-set histograms on screen

Record 02

TESSERA

Thermal-infrared to colour for Landsat imagery, for ISRO's Bharatiya Antariksh Hackathon 2026, with a map of where the model is guessing.

Status
Research
Year
2026
Role
Team leader, team of 4
Input
Landsat ST_B10 (thermal)
PSNR
24.84 dB
Hallucination proxy
0.024 (lower is better)
Time per tile
227 ms

Briefing

Problem
Colourise single-band thermal Landsat tiles so they are easier to interpret, and show where the colour is a guess.
What exists today
A trained two-stage pipeline, evaluated on held-out Landsat test scenes.
Key result
Kept the model with the lower hallucination proxy (0.024) over a distilled student with higher PSNR (26.23 vs 24.84 dB) but a proxy of 0.121.

Objective

Bharatiya Antariksh Hackathon 2026, Problem Statement 10: colourise and enhance infrared satellite imagery so objects are easier to interpret. A single-band thermal tile goes in; a sharpened infrared image, a colour interpretation, and a per-pixel risk map come out. The risk map marks where the colour is not supported by the input.

Architecture

Schematic

  1. Input

    • Landsat ST_B10 thermal tile

  2. Stage 1

    • Infrared enhancement (ResUNet / EDSR-style)

  3. Stage 2

    • ResUNet colouriser
    • Semantic and uncertainty heads

  4. Output

    • Colour image
    • Hallucination-risk map

A teacher that also sees spectral indices (NDVI, NDWI, MNDWI, NDBI) was trained as an upper bound. It is not deployable: the deployed model sees only the thermal band.

Constraints

The challenge allows only the thermal band at inference, so anything else can guide training but never ship. Patches are split by scene and date, so no scene appears in both training and test.

Measurements

Model output, held-out test scenes

Test scene

  1. Input

    ST_B10 thermal, Forest/vegetation proxy (test)

    ST_B10 thermal

  2. Output

    TESSERA RGB, Forest/vegetation proxy (test)

    TESSERA RGB

  3. Risk map

    TESSERA risk, Forest/vegetation proxy (test)

    TESSERA risk

  4. For comparison

    Ground truth RGB, Forest/vegetation proxy (test)

    Ground truth RGB

Forest/vegetation proxy (test) Held-out Landsat test scenes, cropped from the results grid. The model sees only the thermal band; ground truth is shown for comparison. The risk map marks where the colour is not supported by the input.

  1. Input

    ST_B10 thermal, Agriculture (test)

    ST_B10 thermal

  2. Output

    TESSERA RGB, Agriculture (test)

    TESSERA RGB

  3. Risk map

    TESSERA risk, Agriculture (test)

    TESSERA risk

  4. For comparison

    Ground truth RGB, Agriculture (test)

    Ground truth RGB

Agriculture (test) Held-out Landsat test scenes, cropped from the results grid. The model sees only the thermal band; ground truth is shown for comparison. The risk map marks where the colour is not supported by the input.

  1. Input

    ST_B10 thermal, Urban/coastal proxy (test)

    ST_B10 thermal

  2. Output

    TESSERA RGB, Urban/coastal proxy (test)

    TESSERA RGB

  3. Risk map

    TESSERA risk, Urban/coastal proxy (test)

    TESSERA risk

  4. For comparison

    Ground truth RGB, Urban/coastal proxy (test)

    Ground truth RGB

Urban/coastal proxy (test) Held-out Landsat test scenes, cropped from the results grid. The model sees only the thermal band; ground truth is shown for comparison. The risk map marks where the colour is not supported by the input.

  1. Input

    ST_B10 thermal, Water/coast (test)

    ST_B10 thermal

  2. Output

    TESSERA RGB, Water/coast (test)

    TESSERA RGB

  3. Risk map

    TESSERA risk, Water/coast (test)

    TESSERA risk

  4. For comparison

    Ground truth RGB, Water/coast (test)

    Ground truth RGB

Water/coast (test) Held-out Landsat test scenes, cropped from the results grid. The model sees only the thermal band; ground truth is shown for comparison. The risk map marks where the colour is not supported by the input.

  1. Input

    ST_B10 thermal, Desert/bare soil (test)

    ST_B10 thermal

  2. Output

    TESSERA RGB, Desert/bare soil (test)

    TESSERA RGB

  3. Risk map

    TESSERA risk, Desert/bare soil (test)

    TESSERA risk

  4. For comparison

    Ground truth RGB, Desert/bare soil (test)

    Ground truth RGB

Desert/bare soil (test) Held-out Landsat test scenes, cropped from the results grid. The model sees only the thermal band; ground truth is shown for comparison. The risk map marks where the colour is not supported by the input.

Model output, held-out test set

Results grid for five Landsat test scenes: thermal input, baseline colourisation, TESSERA colourisation, true colour, absolute error, and hallucination-risk map.
Held-out test scenes. Columns: thermal input, baseline, TESSERA, ground truth, absolute error, risk.

Exact values. PSNR (peak signal-to-noise ratio, in dB) compares the colour output with the true colour image: higher is better. The hallucination proxy scores colour that the thermal input does not support: lower is better.

PipelinePSNRHallucination proxyTime per tile
TESSERA, deployed24.8357860.023902226.78 ms
With infrared enhancement stage24.8402770.023861299.82 ms
Distilled student (rejected)26.2289750.121392249.57 ms

Failures and iterations

The baseline, trained on pixel loss alone, painted every land cover the same brown. Temperature does not determine colour, so the loss rewarded the average. That failure motivated the separate enhancement stage, the semantic head, and the risk map.

Distilling the teacher into a thermal-only student raised PSNR, but its hallucination proxy was 0.121 against 0.024, so it was rejected. On the pilot split, the largest model tried (8,748,239 parameters) was worse than the baseline on PSNR and much slower.

Status

Research

End of record R-02

Next record R-03 VAJRA

Questions about this record? hello@vedantsomani.tech or open a channel.