Datasets:
Update README.md
Browse files
README.md
CHANGED
@@ -64,7 +64,7 @@ By leveraging these two sensors, the dataset combines **satellite multi-spectral
|
|
64 |
|
65 |
## Taco dataset
|
66 |
|
67 |
-
**Dataset Construction:** We paired **1234** low-resolution (LR) Sentinel-2 images at **10m** and **20m** resolution with high-resolution (HR) NEON images at **2.5m** resolution (resampled from original 1m) that overlap in location and were captured on near dates (often the **same day** or within a few days) to minimize land cover changes. All sites are within the continental United States, where NEON conducts airborne campaigns and Sentinel-2 has routine coverage. Typical NEON flights occur during peak vegetation season with clear conditions, and only cloud-free Sentinel-2 acquisitions (as indicated by quality flags) were selected. Each paired sample covers a **5.
|
68 |
|
69 |
**Spectral Bands:** The dataset includes **12 spectral bands** matching Sentinel-2’s optical bands **B1–B9 and B11–B12** (Band 10 is omitted as explained). For each Sentinel-2 band, the NEON hyperspectral reflectance was spectrally averaged or integrated to simulate the same bandpass. NEON’s fine spectral resolution allows close emulation of Sentinel-2 band responses, including narrow bands like the red-edge (B5–B7 ~705–783 nm) and water vapor band (B9 ~945 nm). The table below summarizes the included bands:
|
70 |
|
@@ -89,7 +89,7 @@ The table below lists each Sentinel-2 band, its approximate central wavelength,
|
|
89 |
|
90 |
> **Table:** Spectral Bands of the Dataset. The central wavelengths and bandwidths correspond to Sentinel-2A. The native 20 m and 60 m bands have been resampled to 10 m to form the Sentinel-2 LR multiband image. “Simulated NEON” indicates that the HR image band was obtained by integrating multiple NEON hyperspectral bands within the equivalent spectral range.
|
91 |
|
92 |
-
> **Note:** *The "idx" column corresponds to the zero-based band index used in the LR/HR image stacks (for the 12 included bands)
|
93 |
|
94 |
|
95 |
### Compression for GeoTIFF Files:
|
@@ -104,9 +104,9 @@ The dataset uses compression settings to reduce storage size while maintaining d
|
|
104 |
|
105 |
Each sample consists of two GeoTIFF files packaged in the **tortilla format**:
|
106 |
|
107 |
-
- **LR image (Sentinel-2):** A single GeoTIFF containing 12 bands at 10 m resolution (excluding B10). Dimensions:
|
108 |
|
109 |
-
- **HR image (NEON):** A GeoTIFF with 12 bands at 2.5 m resolution for the same area and date. Dimensions:
|
110 |
|
111 |
The dataset is organized in **[TACO](https://huggingface.co/datasets/tacofoundation/sen2neon)** multi-part files for direct use with the TACO framework.
|
112 |
|
|
|
64 |
|
65 |
## Taco dataset
|
66 |
|
67 |
+
**Dataset Construction:** We paired **1234** low-resolution (LR) Sentinel-2 images at **10m** and **20m** resolution with high-resolution (HR) NEON images at **2.5m** resolution (resampled from original 1m) that overlap in location and were captured on near dates (often the **same day** or within a few days) to minimize land cover changes. All sites are within the continental United States, where NEON conducts airborne campaigns and Sentinel-2 has routine coverage. Typical NEON flights occur during peak vegetation season with clear conditions, and only cloud-free Sentinel-2 acquisitions (as indicated by quality flags) were selected. Each paired sample covers a **5.16 × 5.16 km** area (5160m × 5160m) – corresponding to a **516×516 px Sentinel-2 tile (10m/px)** and the same area in a **2048×2048 px NEON image (2.5m/px)**. The NEON hyperspectral data (originally 1m) were resampled to 2.5m resolution by averaging blocks of 2×2m (or via appropriate reprojection) to reduce data volume and align to an exact 4:1 scale difference with Sentinel-2. This ensures each Sentinel-2 pixel aligns with a 4×4 block of NEON pixels, simplifying error calculations for super-resolved outputs.
|
68 |
|
69 |
**Spectral Bands:** The dataset includes **12 spectral bands** matching Sentinel-2’s optical bands **B1–B9 and B11–B12** (Band 10 is omitted as explained). For each Sentinel-2 band, the NEON hyperspectral reflectance was spectrally averaged or integrated to simulate the same bandpass. NEON’s fine spectral resolution allows close emulation of Sentinel-2 band responses, including narrow bands like the red-edge (B5–B7 ~705–783 nm) and water vapor band (B9 ~945 nm). The table below summarizes the included bands:
|
70 |
|
|
|
89 |
|
90 |
> **Table:** Spectral Bands of the Dataset. The central wavelengths and bandwidths correspond to Sentinel-2A. The native 20 m and 60 m bands have been resampled to 10 m to form the Sentinel-2 LR multiband image. “Simulated NEON” indicates that the HR image band was obtained by integrating multiple NEON hyperspectral bands within the equivalent spectral range.
|
91 |
|
92 |
+
> **Note:** *The "idx" column corresponds to the zero-based band index used in the LR/HR image stacks (for the 12 included bands).*
|
93 |
|
94 |
|
95 |
### Compression for GeoTIFF Files:
|
|
|
104 |
|
105 |
Each sample consists of two GeoTIFF files packaged in the **tortilla format**:
|
106 |
|
107 |
+
- **LR image (Sentinel-2):** A single GeoTIFF containing 12 bands at 10 m resolution (excluding B10). Dimensions: 516 × 516 pixels × 12 bands.
|
108 |
|
109 |
+
- **HR image (NEON):** A GeoTIFF with 12 bands at 2.5 m resolution for the same area and date. Dimensions: 2064 × 2064 pixels × 12 bands.
|
110 |
|
111 |
The dataset is organized in **[TACO](https://huggingface.co/datasets/tacofoundation/sen2neon)** multi-part files for direct use with the TACO framework.
|
112 |
|