satellogic/EarthView · Observations in the NEON Section of the Dataset (Dates, EPSG Metadata, and the earthview Package)

Hello! First of all, thank you very much for this great contribution to the field of remote sensing. I have been reviewing the NEON section in EarthView (the neon folder) and found a few points that may need clarification or adjustment:

1. Dates and Revisits

The paper mentions multiple annual revisits for each NEON location. However, in the files available on Hugging Face (neon/train and neon/val), each *.h5py file consistently contains only three dates. Checking the script’s output confirms that the repeated dates are generally:

2018-01-01T00:00:00
2019-01-01T00:00:00
2020-01-01T00:00:00 or 2021-01-01T00:00:00

No other intermediate dates or daily/monthly variations that would indicate multiple revisits throughout the year are observed.

import h5py
import glob

train_path = "neon/train"
filenames = glob.glob(train_path + "/**/*.h5py", recursive=True)

for filename in filenames:
    with h5py.File(filename, "r") as f:
        timestamps = f.attrs.get("timestamp", None)
        print(f"File: {filename}")
        print("Timestamps in attrs:", timestamps)

File: neon/train/29_610184661630115N_82_02501461594764W/neon.h5py
Timestamps in attrs: ['2018-01-01T00:00:00' '2019-01-01T00:00:00' '2021-01-01T00:00:00']

File: neon/train/39_11119616886158N_96_620662908988W/neon.h5py
Timestamps in attrs: ['2018-01-01T00:00:00' '2019-01-01T00:00:00' '2020-01-01T00:00:00']

File: neon/train/33_4284794901218N_97_53828320575133W/neon.h5py
Timestamps in attrs: ['2018-01-01T00:00:00' '2019-01-01T00:00:00' '2021-01-01T00:00:00']
...

File: neon/train/47_12316899325486N_99_25287582102743W/neon.h5py
Timestamps in attrs: ['2019-01-01T00:00:00' '2020-01-01T00:00:00' '2021-01-01T00:00:00']

2. NEON Sample Count

According to the paper, there should be a total of 35,501 patches (each having 3 revisits) for NEON-RGB and NEON-Hyper. However, the following code—which uses huggingface_hub to list all files in the repository and filter only those under neon/train/ and neon/val/ with the .h5py extension—reveals a different picture:

from huggingface_hub import list_repo_files

repo_id = "satellogic/EarthView"

# 1) List all files in the repository (type "dataset")
all_files = list_repo_files(repo_id, repo_type="dataset")

# 2) Filter files in neon/train/ or neon/val/ that end with .h5py
train_files = [f for f in all_files if f.startswith("neon/train/") and f.endswith(".h5py")]
val_files  = [f for f in all_files if f.startswith("neon/val/")   and f.endswith(".h5py")]

# 3) Print counts and show some examples
print(f"Total files in neon/train/ with .h5py extension: {len(train_files)}")
print("Examples (train):", train_files[:5], "...\n")

print(f"Total files in neon/val/ with .h5py extension: {len(val_files)}")
print("Examples (val):", val_files[:5], "...")

Total files in neon/train/ with .h5py extension: 128
Examples (train): [
  'neon/train/29_609628957997028N_82_0258062601623W/neon.h5py',
  'neon/train/29_60982689580051N_82_03945061123085W/neon.h5py',
  'neon/train/29_60992047487018N_81_9879842970084W/neon.h5py',
  'neon/train/29_610040512157497N_82_0184201916409W/neon.h5py',
  'neon/train/29_610066039108084N_82_02868168058788W/neon.h5py'
] ...

Total files in neon/val/ with .h5py extension: 334
Examples (val): [
  'neon/val/65_23299667266886N_147_45175544166426W/neon.h5py',
  'neon/val/65_23301059543665N_147_4878112782365W/neon.h5py',
  'neon/val/65_23303191811213N_147_672391950543W/neon.h5py',
  'neon/val/65_23304409602434N_147_55357195318004W/neon.h5py',
  'neon/val/65_23310188697575N_147_47327641858885W/neon.h5py'
] ...

We see that:

There are 128 .h5py files in the neon/train/ directory.
There are 334 .h5py files in the neon/val/ directory.

3. Metadata (coordinates and EPSG)

The bounding box in each *.h5py file appears in lat/lon, while the epsg attribute often indicates a UTM system (e.g., EPSG:32617). It would be useful to clarify if the coordinates are in UTM projection or if there’s a mismatch between bounds and epsg.
The earthview package is mentioned for handling metadata conversion, but it’s unclear if this transformation is implemented elsewhere.

4. earthview Package

The Python package earthview (e.g., ev.load_dataset("neon"), ev.item_to_images("neon")) is referenced, but it’s not available on PyPI or any public repository.
Is there a plan to release it, or is it an internal library?

5. Differences with Paper Description

The paper (Table 1) states RGB images have 0.1 m resolution and hyperspectral images 1 m resolution, which matches the Hugging Face files. However, the total spatial dimension and number of published patches don’t match the reported 35,501 unique locations.
The paper mentions multiple annual revisits, but the files only contain 3 fixed dates.

Final Comments

The dataset looks promising. I’d appreciate clarification on:

How the contents of the neon folder align with the figures in the paper (patch numbers, revisits, etc.).
If there are plans to update the repository or correct inconsistencies (dates, EPSG, etc.).
The status of the earthview package.

Thank you for sharing this valuable resource! Looking forward to updates.