kshitijrajsharma/geobench1-vhr

m-pv4ger (Tortilla Mirror)

This is a format-only repackaging of the GEO-Bench v1 m-pv4ger subset as a single tortilla file readable via tacoreader.v1. The contents are bit-identical to the upstream GEO-Bench v1 release.

This is not a new dataset. It is a container-format change. No samples were added, removed, transformed, normalized, downsampled, or relabelled. The repackaging exists so downstream code can read the data without the geobench pip package, which pins pandas<2.0 and so cannot be installed on Python 3.13+.

What's inside

File	mpv4ger/geobench_mpv4ger.tortilla
Samples	13,812 (train 11,814, val 999, test 999)
Image	RGB aerial, 320 x 320, PNG inside the tortilla (lossless)
Labels	binary: 0 = no solar pv, 1 = solar pv
Class balance	train 10,000 / 1,814; val 846 / 153; test 846 / 153
Size on disk	~788 MB

Schema

The tortilla uses the legacy #y-magic format (so tacoreader.v1 can read it). One sample per row, each row is the PNG itself (tortilla:file_format = "PNG"):

Column	Type	Description
tortilla:id	str	train_000000 to test_000998
tortilla:file_format	str	always "PNG"
tortilla:data_split	str	"train" / "val" / "test"
tortilla:offset	int64	byte offset of the PNG inside the tortilla
tortilla:length	int64	byte length of the PNG
label	int64	0 = no solar pv, 1 = solar pv

Verification

Bit-equivalence to the upstream GEO-Bench v1 release was verified at build time:

• per-split sample counts match
• per-class label counts match exactly
• np.array_equal(...) on 9 random samples (3 per split) returns True for both the pixel arrays and the labels (the PNG decoder produces the same uint8 array that geobench.dataset.GeobenchDataset returns for .bands)

Upstream sources and credit

The data is the m-pv4ger subset of GEO-Bench v1, itself sampled from the PV4GER dataset (Germany aerial RGB with manual solar-PV-panel annotations).

GEO-Bench v1 paper: Lacoste, Alexandre, et al. "GEO-Bench: Toward Foundation Models for Earth Monitoring." NeurIPS Datasets and Benchmarks, 2023. https://arxiv.org/abs/2306.03831

GEO-Bench v1 release: https://github.com/ServiceNow/geo-bench

PV4GER paper: Mayer, Kevin, et al. "3D-PV-Locator: Large-scale detection of rooftop-mounted photovoltaic systems in 3D." Applied Energy 310 (2022): 118469. https://doi.org/10.1016/j.apenergy.2022.118469

Source zip mirror used to build this tortilla: recursix/geo-bench-1.0 at commit 1e5754337032f097396db62cc15a5ff88f4618e8, file classification_v1.0/m-pv4ger.zip.

License

Inherits MIT from the upstream GEO-Bench v1 release.

Loading example

import io
import numpy as np
import tacoreader.v1 as tacoreader
from PIL import Image

df = tacoreader.load("geobench_mpv4ger.tortilla")
train = df[df["tortilla:data_split"] == "train"].reset_index(drop=True)

# A single sample
from pangaea.datasets.utils import read_tortilla_subfile_bytes
png = read_tortilla_subfile_bytes(train.read(0))
arr = np.array(Image.open(io.BytesIO(png)))   # (320, 320, 3) uint8
label = int(train.iloc[0]["label"])           # 0 or 1