heightmap
imagewidth (px)
256
256
latitude
stringclasses
115 values
longitude
stringclasses
329 values
n47
e006
n26
w113
n52
w002
s28
e117
n38
e073
n35
e086
n40
w102
s37
w066
s01
e020
n27
e068
n02
w056
n21
e098
n32
w091
n58
e122
n24
w001
n00
e021
n38
e083
n03
e012
n48
e108
s33
e150
n45
e038
n30
e010
n48
e077
s23
w058
n28
e017
s37
w072
n31
e074
s04
e039
n29
e075
n53
e078
n09
w006
n35
e072
n48
e048
s24
w062
n39
e034
n23
e090
n44
w089
s09
w039
s01
e022
n25
e117
n29
e015
n46
w071
s31
e125
n04
w069
n36
e113
s32
w063
n58
e026
n30
w005
n21
e018
n48
w081
s35
e117
s22
e029
n38
w107
n38
e082
n48
e003
n50
e055
n54
e030
n33
w102
n34
w119
s23
e137
n14
w087
n46
e015
n26
e066
s24
e130
n42
e121
n24
e082
n34
w087
n44
e130
s08
w073
n39
w099
n27
e036
s18
w041
n39
e113
n38
e113
s19
e015
n35
e134
s04
w043
n36
e002
n36
e003
n14
e015
n41
e126
n39
w104
n42
w103
n34
e133
s20
w051
n25
e007
n36
w105
s10
w049
n39
e093
n46
e112
n46
e124
n47
e126
n37
e041
n46
e056
n40
e075
n32
e004
n48
w057
n30
w109
n34
e088
n02
e016

World Heightmaps 256px

This is a dataset of 256x256 Earth heightmaps generated from SRTM 1 Arc-Second Global. Each heightmap is labelled according to its latitude and longitude. There are 573,995 samples. It is the same as World Heightmaps 360px but downsampled to 256x256.

Method

  1. Convert GeoTIFFs into PNGs with Rasterio.
import rasterio
import matplotlib.pyplot as plt
import os

input_directory = '...'
output_directory = '...'
file_list = os.listdir(input_directory)

for i in range(len(file_list)):
    image = rasterio.open(input_directory + file_list[i])
    plt.imsave(output_directory + file_list[i][0:-4] + '.png', image.read(1), cmap='gray')
  1. Split PNGs into 100 patches with Split Image.
from split_image import split_image
import os

input_directory = '...'
output_directory = '...'
file_list = os.listdir(input_directory)

for i in range(len(file_list)):
    split_image(input_directory + file_list[i], 10, 10, should_square=True, should_cleanup=False, output_dir=output_directory)
  1. Hand pick a dataset of corrupted and uncorrupted heightmaps then train a discriminator to automatically filter the whole dataset.

  2. Downsample from 360x360 to 256x256 with Pillow and the Lanczos resampling method.

import glob
from PIL import Image

paths = glob.glob('world-heightmaps-360px-png/data/*/*')

for file_name in paths:
    image = Image.open(file_name)
    if image.width == 256:
        continue
    print(file_name)
    image = image.resize((256, 256), resample=Image.LANCZOS)
    image.save(file_name)
  1. Compile images into parquet files.
import pyarrow as pa
import pyarrow.parquet as pq
import pandas as pd
from PIL import Image
import os
import io
import json

samples_per_file = 10_000

root_dir = 'data/datasets/world-heightmaps-256px-png'
df = pd.read_csv(os.path.join(root_dir, 'metadata.csv'))
df = df.sample(frac=1).reset_index(drop=True)

def save_table(image_data, table_number):
    print(f'Entries in table {table_number}: {len(image_data)}')
    schema = pa.schema(
        fields=[
            ('heightmap', pa.struct([('bytes', pa.binary()), ('path', pa.string())])),
            ('latitude', pa.string()),
            ('longitude', pa.string())
        ],
        metadata={
            b'huggingface': json.dumps({
                'info': {
                    'features': {
                        'heightmap': {'_type': 'Image'},
                        'latitude': {'_type': 'Value', 'dtype': 'string'},
                        'longitude': {'_type': 'Value', 'dtype': 'string'}
                    }
                }
            }).encode('utf-8')
        }
    )

    table = pa.Table.from_pylist(image_data, schema=schema)
    pq.write_table(table, f'data/world-heightmaps-256px-parquet/{str(table_number).zfill(4)}.parquet')

image_data = []
samples_in_current_file = 0
current_file_number = 0
for i, row in df.iterrows():
    if samples_in_current_file >= samples_per_file:
        save_table(image_data, current_file_number)
        image_data = []
        samples_in_current_file = 0
        current_file_number += 1
    samples_in_current_file += 1
    image_path = row['file_name']
    with Image.open(os.path.join(root_dir, image_path)) as image:
        image_bytes = io.BytesIO()
        image.save(image_bytes, format='PNG')
        image_dict = {
            'heightmap': {
                'bytes': image_bytes.getvalue(),
                'path': image_path
            },
            'latitude': str(row['latitude']),
            'longitude': str(row['longitude'])
        }
        image_data.append(image_dict)

save_table(image_data, current_file_number)
Downloads last month
367

Models trained or fine-tuned on novaia/world-heightmaps-256px

Collection including novaia/world-heightmaps-256px