rico-hdl

A fast and easy-to-use remote sensing image format converter for high-throughput deep-learning (rico-hdl).

Overview

The core idea is to run the encoder on a supported remote sensing dataset and use the resulting output to efficiently train deep-learning models. The encoder converts the remote sensing images into a DL-optimized format. The resulting output will provide significantly higher throughput than the original remote sensing images (patches) and should be used instead of the unprocessed dataset. The data is encoded in a DL-framework independent format, ensuring flexible use. Concretely, the image files are converted into the safetensors format and stored inside the LMDB key-value database.

Important

The encoded image data values are identical to the data values from the original dataset!

To access the data with Python, install the LMDB and safetensors packages.

Download

Great care has been taken to ensure the application can effortlessly run on different environments without requiring additional dependencies on the server. To make this possible, the application is packaged in two different ways as an:

To run the application on any x86-64 Linux server, we recommend to use the AppImage:

rico-hdl.AppImage

The docker image can be used to run it on other operating systems:

ghcr.io/kai-tub/rico-hdl:latest

Supported Remote Sensing Datasets

Currently, rico-hdl supports:

BigEarthNet-S1 v2.0
BigEarthNet-S2 v2.0
BigEarthNet-MM v2.0
HySpecNet-11k
UC Merced Land Use
EuroSAT
SSL4EO-S12
Major-Tom-Core

Additional datasets will be added in the near future.

BigEarthNet Example

First, download the rico-hdl binary and install the Python lmdb and saftensors packages. Then, to convert the Sentinel-1 and Sentinel-2 patches from the BigEarthNet v2.0 dataset into the optimized format, call the application with:

rico-hdl bigearthnet --bigearthnet-s1-dir <S1_ROOT_DIR> --bigearthnet-s2-dir <S2_ROOT_DIR> --target-dir Encoded-BigEarthNet

In BigEarthNet, each band is stored as a separate file with the associate band as a suffix (_B01, _B12, _VV, ...). The encoder groups all image files with the same name/prefix and stores the data as a safetensors dictionary, where the dictionary's key is the band name (B01, B12, VV, ...).

Example Input

├── <S1_ROOT_DIR>
│  └── S1A_IW_GRDH_1SDV_20170613T165043
│     └── S1A_IW_GRDH_1SDV_20170613T165043_33UUP_70_48
│        ├── S1A_IW_GRDH_1SDV_20170613T165043_33UUP_70_48_VH.tif
│        └── S1A_IW_GRDH_1SDV_20170613T165043_33UUP_70_48_VV.tif
└── <S2_ROOT_DIR>
   └── S2A_MSIL2A_20170613T101031_N9999_R022_T33UUP
      └── S2A_MSIL2A_20170613T101031_N9999_R022_T33UUP_75_43
         ├── S2A_MSIL2A_20170613T101031_N9999_R022_T33UUP_75_43_B01.tif
         ├── S2A_MSIL2A_20170613T101031_N9999_R022_T33UUP_75_43_B02.tif
         ├── S2A_MSIL2A_20170613T101031_N9999_R022_T33UUP_75_43_B03.tif
         ├── S2A_MSIL2A_20170613T101031_N9999_R022_T33UUP_75_43_B04.tif
         ├── S2A_MSIL2A_20170613T101031_N9999_R022_T33UUP_75_43_B05.tif
         ├── S2A_MSIL2A_20170613T101031_N9999_R022_T33UUP_75_43_B06.tif
         ├── S2A_MSIL2A_20170613T101031_N9999_R022_T33UUP_75_43_B07.tif
         ├── S2A_MSIL2A_20170613T101031_N9999_R022_T33UUP_75_43_B08.tif
         ├── S2A_MSIL2A_20170613T101031_N9999_R022_T33UUP_75_43_B09.tif
         ├── S2A_MSIL2A_20170613T101031_N9999_R022_T33UUP_75_43_B8A.tif
         ├── S2A_MSIL2A_20170613T101031_N9999_R022_T33UUP_75_43_B11.tif
         └── S2A_MSIL2A_20170613T101031_N9999_R022_T33UUP_75_43_B12.tif

LMDB Result

'S1A_IW_GRDH_1SDV_20170613T165043_33UUP_70_48':
  {
    'VH': <120x120 float32 safetensors image data>
    'VV': <120x120 float32 safetensors image data>
  },
'S2A_MSIL2A_20170613T101031_N9999_R022_T33UUP':
  {
    'B01': <30x30   uint16 safetensors image data>,
    'B02': <120x120 uint16 safetensors image data>,
    'B03': <120x120 uint16 safetensors image data>,
    'B04': <120x120 uint16 safetensors image data>,
    'B05': <60x60   uint16 safetensors image data>,
    'B06': <60x60   uint16 safetensors image data>,
    'B07': <60x60   uint16 safetensors image data>,
    'B08': <120x120 uint16 safetensors image data>,
    'B8A': <60x60   uint16 safetensors image data>,
    'B09': <30x30   uint16 safetensors image data>,
    'B11': <60x60   uint16 safetensors image data>,
    'B12': <60x60   uint16 safetensors image data>,
  }

The following code shows how to access the converted database:

import lmdb
# import desired deep-learning library:
# numpy, torch, tensorflow, paddle, flax, mlx
from safetensors.numpy import load
from pathlib import Path

# path to the encoded dataset/output of rico-hdl
encoded_path = Path("./Encoded-BigEarthNet")

# Make sure to only open the environment once
# and not everytime an item is accessed.
env = lmdb.open(str(encoded_path), readonly=True)

with env.begin() as txn:
  # string encoding is required to map the string to an LMDB key
  safetensor_dict = load(txn.get("S2A_MSIL2A_20170613T101031_N9999_R022_T33UUP".encode()))

rgb_bands = ["B04", "B03", "B02"]
rgb_tensor = np.stack([safetensor_dict[b] for b in rgb_bands])
assert rgb_tensor.shape == (3, 120, 120)

Tip

Remember to use the appropriate load function for a given deep-learning library.

The ConfigILM library provides an LMDB reader example that shows how to utilize the encoded data for high-throughput deep-learning.

HySpecNet-11k Example

First, download the rico-hdl binary and install the Python lmdb and saftensors packages. Then, to convert the patches from the HySpecNet-11k dataset into the optimized format, call the application with:

rico-hdl hyspecnet-11k --dataset-dir <HYSPECNET_ROOT_DIR> --dataset-dir Encoded-HySpecNet

In HySpecNet-11k, each patch contains 224 bands. The encoder will convert each patch into a safetensors dictionary, where the band index prefixed with B is the key (for example, B1, B201) of the safetensor dictionary.

Example Input

integration_tests/tiffs/HySpecNet-11k
├── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X03110438
│  ├── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X03110438-QL_PIXELMASK.TIF
│  ├── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X03110438-QL_QUALITY_CIRRUS.TIF
│  ├── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X03110438-QL_QUALITY_CLASSES.TIF
│  ├── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X03110438-QL_QUALITY_CLOUD.TIF
│  ├── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X03110438-QL_QUALITY_CLOUDSHADOW.TIF
│  ├── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X03110438-QL_QUALITY_HAZE.TIF
│  ├── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X03110438-QL_QUALITY_SNOW.TIF
│  ├── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X03110438-QL_QUALITY_TESTFLAGS.TIF
│  ├── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X03110438-QL_SWIR.TIF
│  ├── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X03110438-QL_VNIR.TIF
│  ├── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X03110438-SPECTRAL_IMAGE.TIF
│  └── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X03110438-THUMBNAIL.jpg
└── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X04390566
   ├── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X04390566-QL_PIXELMASK.TIF
   ├── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X04390566-QL_QUALITY_CIRRUS.TIF
   ├── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X04390566-QL_QUALITY_CLASSES.TIF
   ├── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X04390566-QL_QUALITY_CLOUD.TIF
   ├── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X04390566-QL_QUALITY_CLOUDSHADOW.TIF
   ├── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X04390566-QL_QUALITY_HAZE.TIF
   ├── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X04390566-QL_QUALITY_SNOW.TIF
   ├── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X04390566-QL_QUALITY_TESTFLAGS.TIF
   ├── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X04390566-QL_SWIR.TIF
   ├── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X04390566-QL_VNIR.TIF
   ├── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X04390566-SPECTRAL_IMAGE.TIF
   └── ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X04390566-THUMBNAIL.jpg

LMDB Result

[!INFO] The encoder will only process the image data (SPECTRAL_IMAGE.TIF) and skip over the quality indicator and thumbnail files.

'ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X03110438':
  {
    'B1': <128x128 int16 safetensors image data>
    'B2': <128x128 int16 safetensors image data>
     ⋮
    'B10': <128x128 int16 safetensors image data>
    'B11': <128x128 int16 safetensors image data>
     ⋮
    'B100': <128x128 int16 safetensors image data>
    'B101': <128x128 int16 safetensors image data>
     ⋮
    'B223': <128x128 int16 safetensors image data>
    'B224': <128x128 int16 safetensors image data>
  },
'ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X04390566':
  {
    'B1': <128x128 int16 safetensors image data>
    'B2': <128x128 int16 safetensors image data>
     ⋮
    'B10': <128x128 int16 safetensors image data>
    'B11': <128x128 int16 safetensors image data>
     ⋮
    'B100': <128x128 int16 safetensors image data>
    'B101': <128x128 int16 safetensors image data>
     ⋮
    'B223': <128x128 int16 safetensors image data>
    'B224': <128x128 int16 safetensors image data>
  }

import lmdb
import numpy as np
# import desired deep-learning library:
# numpy, torch, tensorflow, paddle, flax, mlx
from safetensors.numpy import load
from pathlib import Path

encoded_path = "Encoded-HySpecNet"

# Make sure to only open the environment once
# and not everytime an item is accessed.
env = lmdb.open(str(encoded_path), readonly=True)

with env.begin() as txn:
  # string encoding is required to map the string to an LMDB key
  safetensor_dict = load(txn.get("ENMAP01-____L2A-DT0000004950_20221103T162438Z_001_V010110_20221118T145147Z-Y01460273_X04390566".encode()))

hyspecnet_bands = range(1, 225)
# recommendation from HySpecNet-11k paper
skip_bands = [126, 127, 128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 160, 161, 162, 163, 164, 165, 166]
tensor = np.stack([safetensor_dict[f"B{k}"] for k in hyspecnet_bands if k not in skip_bands])
assert tensor.shape == (202, 128, 128)

UC Merced Land Use Example

First, download the rico-hdl binary and install the Python lmdb and saftensors packages. Then, to convert the patches from the UC Merced Land Use dataset into the optimized format, call the application with:

rico-hdl uc-merced --dataset-dir <UC_MERCED_LAND_USE_ROOT_DIR> --dataset-dir Encoded-UC-Merced

In UC Merced, each patch contains 3 bands (RGB). The encoder will convert each patch into a safetensors dictionary, where the band's color interpretation is the key (one of Red, Green, Blue) of the safetensor dictionary.

Example Input

integration_tests/tiffs/UCMerced_LandUse
└── Images
   ├── airplane
   │  ├── airplane00.tif
   │  └── airplane42.tif
   └── forest
      ├── forest10.tif
      └── forest99.tif

LMDB Result

'airplane00':
  {
    'Red':   <256x256 uint8 safetensors image data>
    'Green': <256x256 uint8 safetensors image data>
    'Blue':  <256x256 uint8 safetensors image data>
  },
'airplane42':
  {
    'Red':   <256x256 uint8 safetensors image data>
    'Green': <256x256 uint8 safetensors image data>
    'Blue':  <256x256 uint8 safetensors image data>
  },
'forest10':
  {
    'Red':   <256x256 uint8 safetensors image data>
    'Green': <256x256 uint8 safetensors image data>
    'Blue':  <256x256 uint8 safetensors image data>
  },
'forest99':
  {
    'Red':   <256x256 uint8 safetensors image data>
    'Green': <256x256 uint8 safetensors image data>
    'Blue':  <256x256 uint8 safetensors image data>
  }

import lmdb
import numpy as np
# import desired deep-learning library:
# numpy, torch, tensorflow, paddle, flax, mlx
from safetensors.numpy import load
from pathlib import Path

encoded_path = "Encoded-UC-Merced"

# Make sure to only open the environment once
# and not everytime an item is accessed.
env = lmdb.open(str(encoded_path), readonly=True)

with env.begin() as txn:
  # string encoding is required to map the string to an LMDB key
  safetensor_dict = load(txn.get("airplane00".encode()))

tensor = np.stack([safetensor_dict[key] for key in ["Red", "Green", "Blue"]])
assert tensor.shape == (3, 256, 256)

EuroSAT Example

First, download the rico-hdl binary and install the Python lmdb and saftensors packages. Then, to convert the patches from the EuroSAT multi-spectral dataset into the optimized format, call the application with:

rico-hdl eurosat-multi-spectral --dataset-dir <EURO_SAT_MS_ROOT_DIR> --dataset-dir Encoded-EuroSAT-MS

In EuroSAT, each patch contains 13 bands from a Sentinel-2 L1C tile. The encoder will convert each patch into a safetensors where the dictionary's key is the band name (B01, B02,..., B10, B11, B12, B08A) of the safetensor dictionary.

Example Input

integration_tests/tiffs/EuroSAT_MS
├── AnnualCrop
│  └── AnnualCrop_1.tif
├── Pasture
│  └── Pasture_300.tif
└── SeaLake
   └── SeaLake_3000.tif

LMDB Result

'AnnualCrop_1':
  {
    'B01':   <64x64 uint16 safetensors image data>,
    'B02':   <64x64 uint16 safetensors image data>,
    'B03':   <64x64 uint16 safetensors image data>,
    'B04':   <64x64 uint16 safetensors image data>,
    'B05':   <64x64 uint16 safetensors image data>,
    'B06':   <64x64 uint16 safetensors image data>,
    'B07':   <64x64 uint16 safetensors image data>,
    'B08':   <64x64 uint16 safetensors image data>,
    'B09':   <64x64 uint16 safetensors image data>,
    'B10':   <64x64 uint16 safetensors image data>,
    'B11':   <64x64 uint16 safetensors image data>,
    'B12':   <64x64 uint16 safetensors image data>,
    'B08A':  <64x64 uint16 safetensors image data>,
  },
'Pasture_300':
  {
    'B01':   <64x64 uint16 safetensors image data>,
    'B02':   <64x64 uint16 safetensors image data>,
    'B03':   <64x64 uint16 safetensors image data>,
    'B04':   <64x64 uint16 safetensors image data>,
    'B05':   <64x64 uint16 safetensors image data>,
    'B06':   <64x64 uint16 safetensors image data>,
    'B07':   <64x64 uint16 safetensors image data>,
    'B08':   <64x64 uint16 safetensors image data>,
    'B09':   <64x64 uint16 safetensors image data>,
    'B10':   <64x64 uint16 safetensors image data>,
    'B11':   <64x64 uint16 safetensors image data>,
    'B12':   <64x64 uint16 safetensors image data>,
    'B08A':  <64x64 uint16 safetensors image data>,
  },
'SeaLake_3000':
  {
    'B01':   <64x64 uint16 safetensors image data>,
    'B02':   <64x64 uint16 safetensors image data>,
    'B03':   <64x64 uint16 safetensors image data>,
    'B04':   <64x64 uint16 safetensors image data>,
    'B05':   <64x64 uint16 safetensors image data>,
    'B06':   <64x64 uint16 safetensors image data>,
    'B07':   <64x64 uint16 safetensors image data>,
    'B08':   <64x64 uint16 safetensors image data>,
    'B09':   <64x64 uint16 safetensors image data>,
    'B10':   <64x64 uint16 safetensors image data>,
    'B11':   <64x64 uint16 safetensors image data>,
    'B12':   <64x64 uint16 safetensors image data>,
    'B08A':  <64x64 uint16 safetensors image data>,
  }

import lmdb
import numpy as np
# import desired deep-learning library:
# numpy, torch, tensorflow, paddle, flax, mlx
from safetensors.numpy import load
from pathlib import Path

encoded_path = "Encoded-EuroSAT-MS"

# Make sure to only open the environment once
# and not everytime an item is accessed.
env = lmdb.open(str(encoded_path), readonly=True)

with env.begin() as txn:
  # string encoding is required to map the string to an LMDB key
  safetensor_dict = load(txn.get("AnnualCrop_1".encode()))

tensor = np.stack([safetensor_dict[key] for key in [
  "B01",
  "B02",
  "B03",
  "B04",
  "B05",
  "B06",
  "B07",
  "B08",
  "B09",
  "B10",
  "B11",
  "B12",
  "B08A"
]])
assert tensor.shape == (13, 64, 64)

SSL4EO-S12 Example

First, download the rico-hdl binary and install the Python lmdb and saftensors packages. Then, to convert the Sentinel-1, Sentinel-2 L1C, and Sentinel-2 L2A patches from the SSL4EO-S12 dataset into the optimized format, call the application with:

rico-hdl ssl4eo-s12 --s1-dir <S1_ROOT_DIR> --s2-l1c-dir <S2_L1C_ROOT_DIR> --s2-l2a-dir <S2_L2A_ROOT_DIR> --target-dir Encoded-SSL4EO-S12

In SSL4EO-S12, each band is stored as a separate file with the associate band as a name (B1.tif, B9.tif, B10.tif, VV.tif, ...). The encoder groups all image files with the same name/prefix and stores the data as a safetensors dictionary, where the dictionary's key is the band name (B1, B9, B10, VV, ...).

Example Input

<SSL4EO-S12 ROOT DIRECTORY>
├── s1
│  └── 0000200
│     ├── S1A_IW_GRDH_1SDV_20200607T010800_20200607T010825_032904_03CFBA_D457
│     │  ├── metadata.json
│     │  ├── VH.tif
│     │  └── VV.tif
│     └── S1A_IW_GRDH_1SDV_20200903T131212_20200903T131237_034195_03F8F5_AC1C
│        ├── metadata.json
│        ├── VH.tif
│        └── VV.tif
├── s2a
│  └── 0000200
│     ├── 20200604T054639_20200604T054831_T43RCP
│     │  ├── B1.tif
│     │  ├── B2.tif
│     │  ├── B3.tif
│     │  ├── B4.tif
│     │  ├── B5.tif
│     │  ├── B6.tif
│     │  ├── B7.tif
│     │  ├── B8.tif
│     │  ├── B8A.tif
│     │  ├── B9.tif
│     │  ├── B11.tif
│     │  ├── B12.tif
│     │  └── metadata.json
│     └── 20200813T054639_20200813T054952_T43RCP
│        ├── B1.tif
│        ├── B2.tif
│        ├── B3.tif
│        ├── B4.tif
│        ├── B5.tif
│        ├── B6.tif
│        ├── B7.tif
│        ├── B8.tif
│        ├── B8A.tif
│        ├── B9.tif
│        ├── B11.tif
│        ├── B12.tif
│        └── metadata.json
└── s2c
   └── 0000200
      ├── 20200604T054639_20200604T054831_T43RCP
      │  ├── B1.tif
      │  ├── B2.tif
      │  ├── B3.tif
      │  ├── B4.tif
      │  ├── B5.tif
      │  ├── B6.tif
      │  ├── B7.tif
      │  ├── B8.tif
      │  ├── B8A.tif
      │  ├── B9.tif
      │  ├── B10.tif
      │  ├── B11.tif
      │  ├── B12.tif
      │  └── metadata.json
      └── 20200823T054639_20200823T055618_T43RCP
         ├── B1.tif
         ├── B2.tif
         ├── B3.tif
         ├── B4.tif
         ├── B5.tif
         ├── B6.tif
         ├── B7.tif
         ├── B8.tif
         ├── B8A.tif
         ├── B9.tif
         ├── B10.tif
         ├── B11.tif
         ├── B12.tif
         └── metadata.json

LMDB Result

Note: We merge the patch directory with the two upper parent directories. This path merging ensures that values are unique and that the entire SSL4EO-S12 dataset can be stored in a single LMDB database.

And the authors of SSL4EO-S12 did not ensure that the resulting patches have a consistent size! There are some patches that have an additional row/column

's1_0000200_S1A_IW_GRDH_1SDV_20200607T010800_20200607T010825_032904_03CFBA_D457':
  {
    'VH': <264x264 float32 safetensors image data>
    'VV': <264x264 float32 safetensors image data>
  },
's1_0000200_S1A_IW_GRDH_1SDV_20200903T131212_20200903T131237_034195_03F8F5_AC1C':
  {
    'VH': <264x264 float32 safetensors image data>
    'VV': <264x264 float32 safetensors image data>
  },
's2a_0000200_20200604T054639_20200604T054831_T43RCP': {
    'B1':  <44x44   uint16 safetensors image data>
    'B2':  <264x264 uint16 safetensors image data>
    'B3':  <264x264 uint16 safetensors image data>
    'B4':  <264x264 uint16 safetensors image data>
    'B5':  <132x132 uint16 safetensors image data>
    'B6':  <132x132 uint16 safetensors image data>
    'B7':  <132x132 uint16 safetensors image data>
    'B8':  <132x132 uint16 safetensors image data>
    'B8A': <132x132 uint16 safetensors image data>
    'B9':  <44x44   uint16 safetensors image data>
    'B10': <44x44   uint16 safetensors image data>
    'B11': <132x132 uint16 safetensors image data>
    'B12': <132x132 uint16 safetensors image data>
  },
's2a_0000200_20200813T054639_20200813T054952_T43RCP': {
    'B1':  <44x44   uint16 safetensors image data>
    'B2':  <264x264 uint16 safetensors image data>
    'B3':  <264x264 uint16 safetensors image data>
    'B4':  <264x264 uint16 safetensors image data>
    'B5':  <132x132 uint16 safetensors image data>
    'B6':  <132x132 uint16 safetensors image data>
    'B7':  <132x132 uint16 safetensors image data>
    'B8':  <132x132 uint16 safetensors image data>
    'B8A': <132x132 uint16 safetensors image data>
    'B9':  <44x44   uint16 safetensors image data>
    'B10': <44x44   uint16 safetensors image data>
    'B11': <132x132 uint16 safetensors image data>
    'B12': <132x132 uint16 safetensors image data>
  },
's2c_0000200_20200604T054639_20200604T054831_T43RCP': {
    'B1':  <44x44   uint16 safetensors image data>
    'B2':  <264x264 uint16 safetensors image data>
    'B3':  <264x264 uint16 safetensors image data>
    'B4':  <264x264 uint16 safetensors image data>
    'B5':  <132x132 uint16 safetensors image data>
    'B6':  <132x132 uint16 safetensors image data>
    'B7':  <132x132 uint16 safetensors image data>
    'B8':  <132x132 uint16 safetensors image data>
    'B8A': <132x132 uint16 safetensors image data>
    'B9':  <44x44   uint16 safetensors image data>
    'B11': <132x132 uint16 safetensors image data>
    'B12': <132x132 uint16 safetensors image data>
  },
's2c_0000200_20200823T054639_20200823T055618_T43RCP': {
    'B1':  <44x44   uint16 safetensors image data>
    'B2':  <264x264 uint16 safetensors image data>
    'B3':  <264x264 uint16 safetensors image data>
    'B4':  <264x264 uint16 safetensors image data>
    'B5':  <132x132 uint16 safetensors image data>
    'B6':  <132x132 uint16 safetensors image data>
    'B7':  <132x132 uint16 safetensors image data>
    'B8':  <132x132 uint16 safetensors image data>
    'B8A': <132x132 uint16 safetensors image data>
    'B9':  <44x44   uint16 safetensors image data>
    'B11': <132x132 uint16 safetensors image data>
    'B12': <132x132 uint16 safetensors image data>
  },

The following code shows how to access the converted database:

import lmdb
# import desired deep-learning library:
# numpy, torch, tensorflow, paddle, flax, mlx
from safetensors.numpy import load
from pathlib import Path

# path to the encoded dataset/output of rico-hdl
encoded_path = Path("./Encoded-SSL4EO-S12")

# Make sure to only open the environment once
# and not everytime an item is accessed.
env = lmdb.open(str(encoded_path), readonly=True)

with env.begin() as txn:
  # string encoding is required to map the string to an LMDB key
  safetensor_dict = load(txn.get("s2c_0000200_20200823T054639_20200823T055618_T43RCP".encode()))

rgb_bands = ["B4", "B3", "B2"]
rgb_tensor = np.stack([safetensor_dict[b] for b in rgb_bands])
assert rgb_tensor.shape == (3, 264, 264)

Major-TOM-Core Example

First, download the rico-hdl binary and install the Python lmdb and saftensors packages. Then, to convert the Sentinel-1 and Sentinel-2 patches from the Major-TOM-Core dataset into the optimized format, call the application with:

rico-hdl major-tom-core --s1-dir <S1_ROOT_DIR> --s2-dir <S2_ROOT_DIR> --target-dir encoded-major-tom

In Major-TOM-Core, each band is stored as a separate file with the associate band as the name (B01.tif, B12.tif, vv.tif, ...). The directory that contains the bands is the associated product id/patch and is uniquely identifiable if it is combined with the associated grid cell id (parent directory). The encoder groups all unique patches (<grid_cell>_<product_id>) and stores the data as a safetensors dictionary, where the dictionary's key is the band name (B01, B12, vv, ...).

Note

The encoder will not encode the thumbnail.png nor the cloud_mask.tif band!

Example Input

├── <S1_ROOT_DIR>
│  └── 897U
│     └── 897U_171R
│        └── S1B_IW_GRDH_1SDV_20210827T012624_20210827T012653_028425_036437_rtc
│           ├── thumbnail.png
│           ├── vh.tif
│           └── vv.tif
└── <S2_ROOT_DIR>
   └── 199U
      └── 199U_1099R
         └── S2B_MSIL2A_20200223T032739_N9999_R018_T48QUE_20230924T183543
            ├── B01.tif
            ├── B02.tif
            ├── B03.tif
            ├── B04.tif
            ├── B05.tif
            ├── B06.tif
            ├── B07.tif
            ├── B08.tif
            ├── B09.tif
            ├── B8A.tif
            ├── B11.tif
            ├── B12.tif
            ├── cloud_mask.tif
            └── thumbnail.png

LMDB Result

Note: We merge the patch directory with the parent directories. This path merging ensures that values are unique.

And the authors of Major-Tom-Core did not ensure that the resulting patches have a consistent size! There are some patches that have a different size, like 195D_241R/S1B_IW_GRDH_1SDV_20200419T165643_20200419T165708_021215_028426_rtc/vv.tif with a pixel size of (1424, 1424) instead of (1068, 1424) and where each pixel is 7.5x7.5 m instead of 10x10 m (most likely the patch has been accidentally interpolated).

'897U_171R_S1B_IW_GRDH_1SDV_20210827T012624_20210827T012653_028425_036437_rtc':
  {
    'vh': <1068x1068 float32 safetensors image data>
    'vv': <1068x1068 float32 safetensors image data>
  },
'199U_1099R_S2A_MSIL2A_20170613T101031_N9999_R022_T33UUP':
  {
    'B01': <178x178   uint16 safetensors image data>,
    'B02': <1068x1068 uint16 safetensors image data>,
    'B03': <1068x1068 uint16 safetensors image data>,
    'B04': <1068x1068 uint16 safetensors image data>,
    'B05': <534x534   uint16 safetensors image data>,
    'B06': <534x534   uint16 safetensors image data>,
    'B07': <534x534   uint16 safetensors image data>,
    'B08': <1068x1068 uint16 safetensors image data>,
    'B8A': <534x534   uint16 safetensors image data>,
    'B09': <178x178   uint16 safetensors image data>,
    'B11': <534x534   uint16 safetensors image data>,
    'B12': <534x534   uint16 safetensors image data>,
  }

The following code shows how to access the converted database:

import lmdb
# import desired deep-learning library:
# numpy, torch, tensorflow, paddle, flax, mlx
from safetensors.numpy import load
from pathlib import Path

# path to the encoded dataset/output of rico-hdl
encoded_path = Path("./encoded-major-tom")

# Make sure to only open the environment once
# and not everytime an item is accessed.
env = lmdb.open(str(encoded_path), readonly=True)

with env.begin() as txn:
  # string encoding is required to map the string to an LMDB key
  safetensor_dict = load(txn.get("199U_1099R_S2A_MSIL2A_20170613T101031_N9999_R022_T33UUP".encode()))

rgb_bands = ["B04", "B03", "B02"]
rgb_tensor = np.stack([safetensor_dict[b] for b in rgb_bands])
assert rgb_tensor.shape == (3, 1068, 1068)

Tip

Remember to use the appropriate load function for a given deep-learning library.

Design

Why safetensors?

The main advantage of the safetensors format is its fast and deep-learning independent tensor serialization capability. This allows teams with different deep-learning framework preferences to utilize the same data without issues. Please refer to the official documentation to discover more benefits of the safetensors format.

Why LMDB?

LMDB is an in-memory key-value store known for its reliability and high performance. It effectively utilizes the operating system's buffer cache and allows seamless parallel read access. These properties make it an excellent choice for environments where multiple users require access to the same data, which is common in deep-learning research.

One significant advantage of choosing LMDB over more array-structured solutions like netcdf or Zarr is that it is better aligned with the access patterns and dataset characteristics specific to remote sensing datasets for deep-learning. Remote sensing deep-learning datasets typically consist of small images (usually around 120px x 120px) with varying resolutions based on the selected band (e.g., BigEarthNet's highest resolution is 120px x 120px and the lowest is 20px x 20px). These images are randomly accessed during training, which differs from the access patterns in classical machine-learning applications or applications that calculate zonal statistics. These characteristics make array-structured data formats less suitable for deep-learning applications.

Citation

If you use this work, please cite:

@article{clasen2024refinedbigearthnet,
  title={reBEN: Refined BigEarthNet Dataset for Remote Sensing Image Analysis},
  author={Clasen, Kai Norman and Hackel, Leonard and Burgert, Tom and Sumbul, Gencer and Demir, Beg{\"u}m and Markl, Volker},
  year={2024},
  eprint={2407.03653},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2407.03653},
}

Name		Name	Last commit message	Last commit date
Latest commit History 30 Commits
.github/workflows		.github/workflows
deprecated_rust		deprecated_rust
integration_tests		integration_tests
rico_hdl		rico_hdl
.envrc		.envrc
.gitignore		.gitignore
LICENSE		LICENSE
README.md		README.md
flake.lock		flake.lock
flake.nix		flake.nix
poetry.lock		poetry.lock
pyproject.toml		pyproject.toml

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Repository files navigation

rico-hdl

Overview

Download

Supported Remote Sensing Datasets

BigEarthNet Example

HySpecNet-11k Example

UC Merced Land Use Example

EuroSAT Example

SSL4EO-S12 Example

Major-TOM-Core Example

Design

Citation

About

Releases 9

Packages

Languages

License

kai-tub/rico-hdl

Folders and files

Latest commit

History

Repository files navigation

rico-hdl

Overview

Download

Supported Remote Sensing Datasets

BigEarthNet Example

HySpecNet-11k Example

UC Merced Land Use Example

EuroSAT Example

SSL4EO-S12 Example

Major-TOM-Core Example

Design

Citation

About

Topics

Resources

License

Stars

Watchers

Forks

Releases 9

Packages 0

Languages

Packages