Portfolio

Research · 2026

Reading landfills off a satellite

My master's thesis at Politecnico di Milano, inside PERIVALLON. Reading the same frozen classifier at a larger input doubles where it can point on satellite imagery: recall at 10 percent of area goes from 0.282 to 0.560 at 120 cm.

Python · PyTorch · Swin Transformer · Remote Sensing · Weakly Supervised Learning · Experimental Design

2x
recall at 120 cm

same frozen weights, larger input

70
logged experiments

prediction registered before measurement

0.41
held-out error of the readout law

in doublings, no generalisation gap

At 120 cm and a 176 pixel input, the activation map is a 6 by 6 grid

The annotations are image-level labels, so where comes out of the classifier's activation map, whose resolution the network stride sets. A landfill covers part of one cell, and the map cannot point at it even when the classifier is confident. Enlarging the input at inference makes the grid finer without touching a weight, at a compute cost quadratic in the side.

The readout law is cell = max(0.47 * object_size, 5.5 * GSD), fitted on 26,041 objects

I fitted the law on per-object observations from five sources: it sets the output cell for a given object size and ground sampling distance, and its error is 0.409 doublings on the fit against 0.411 on held-out aerial data.

The readout that moves the pointing game 0.080 to 0.307 drops AUROC 0.790 to 0.719

Both at 120 cm, six seeds per readout size: the change that fixes pointing hurts deciding. At 30 cm the classifier falls to chance, AUROC 0.581 and F1 0.000, while the map still places 71.6 percent of annotated sites inside the hottest 10 percent of area. On a second dataset and a second sensor, aerial orthophotos, AUROC moves minus 17.8 percent and the pointing game plus 148.8 percent, bootstrap interval minus 0.210 to minus 0.144.

At 1 percent of territory visited the enlarged readout finds about 2.5 times as much at 120 cm

An inspector asks what share of sites a given amount of territory visited turns up. At 5 percent visited:

Ground resolutionNative readoutEnlarged readout
120 cm0.2430.463
60 cm0.5500.736
30 cm0.7090.762

A model beat the oracle of its own grid on 2026-07-25, and the rewrite found three defects

Every experiment registers its prediction before the measurement, and the log keeps both. That impossible score sent me back through the localisation evaluation, where one of three defects biased the IoU by an amount that depended on grid size, and grid size is what the thesis compares. Only post-rewrite numbers appear here, and a per-object ceiling of 0.772 was retracted afterwards as a best-of-four that chance beats.

Validation rank predicts test rank at Spearman 0.966 across 44 configurations

On recall at 10 percent of area, the configuration picked on validation is also the best on test. The work sits inside PERIVALLON, a Horizon Europe project; the repository is private and the thesis is in writing for December 2026.