IntegraCAR-LULC-10K: A High-Resolution Optical Satellite Dataset for LULC Segmentation in the Brazilian Rural Environmental Registry

⚠️ High Volume & Storage Advisory (+300 GB)

This repository hosts the complete 10,000-tile collection (IntegraCAR-LULC-10K), consisting of over 300 GB of high-resolution satellite imagery ( $2048\times 20482048\; \backslash times\; 2048$ px at $0.5\text{ m/px}0.5\backslash text\{\; m/px\}$ ) and pixel-level segmentation masks stored in Parquet shards.

• Prototyping & Quickstart: If you do not have sufficient disk storage or wish to perform initial validation and quick prototyping, we strongly recommend using the lightweight benchmark subset laicsiifes/IntegraCAR-LULC-500 (~10–15 GB), which preserves the exact same 10-region geographic stratification and 5-class taxonomy.
• Streaming Mode: You can stream samples on-the-fly without downloading the entire 300 GB dataset to your local drive by setting streaming=True (see Usage Examples below).
• Local Download: If you plan to download the full dataset offline, ensure you have at least 350–400 GB of free high-speed storage (NVMe SSD recommended) and configure your cache directory accordingly.

Dataset Summary

IntegraCAR-LULC-10K is a large-scale, high-resolution remote sensing dataset designed for Land-Use and Land-Cover (LULC) semantic segmentation in the context of the Brazilian Rural Environmental Registry (Cadastro Ambiental Rural - CAR). Developed by the Laboratório de Inteligência Artificial e Ciência de Dados (LAICSI) at the Instituto Federal do Espírito Santo (IFES) in collaboration with the Instituto de Defesa Agropecuária e Florestal do Espírito Santo (IDAF), this dataset is presented in conjunction with our SIBGRAPI 2026 publication:

IntegraCAR-LULC-10K: A High-Resolution Optical Satellite Dataset for LULC Segmentation in the Brazilian Rural Environmental Registry
Calebe Albertino, Gabriel M. B. Lima, Vinícius R. Oliveira, Arthur R. V. Ribeiro, Eliza K. S. Oliveira, Eduardo H. P. Souza, Otavio G. Dalvi, Karin S. Komati, Jefferson O. Andrade, Francisco A. Boldt, and Thiago M. Paixão. (SIBGRAPI 2026)

The full collection comprises 10,000 optical satellite image tiles ( $2048\times 20482048\; \backslash times\; 2048$ pixels, $0.5\text{ m/pixel}0.5\backslash text\{\; m/pixel\}$ spatial resolution) covering $1,024\text{ m}\times 1,024\text{ m}1,024\backslash text\{\; m\}\; \backslash times\; 1,024\backslash text\{\; m\}$ on the ground per tile, each paired with dense pixel-level thematic annotations across five CAR-aligned classes.

---

Hardware & Storage Recommendations

Because the full dataset comprises over 300 GB across several Parquet shards, consider the following recommendations before starting:

Requirement	Recommended Specification	Notes
Free Disk Space	$\ge \backslash ge$ 400 GB	Required if doing a full local download. Prefer fast NVMe/SSD storage to avoid I/O bottlenecks.
System RAM	$\ge \backslash ge$ 32 GB	Prevents out-of-memory errors when uncompressing large batches of $2048\times 20482048\; \backslash times\; 2048$ tiles.
Cache Management	Custom HF_HOME	By default, Hugging Face caches in ~/.cache/huggingface. Point this to a high-capacity drive before loading.
Alternative Method	Streaming (streaming=True)	Streams tiles on-the-fly directly from the Hugging Face Hub with zero disk overhead.

Redirecting the Hugging Face Cache:

# Linux / macOS
export HF_HOME="/path/to/large_storage/huggingface"

# Windows (PowerShell)
$env:HF_HOME="D:\path\to\large_storage\huggingface"

---

Data Acquisition & Extraction Pipeline

Data acquisition is based on publicly available geospatial layers from the GeoBases platform (https://geobases.es.gov.br/), reflecting orthophotomosaics and thematic maps created by specialists from the Jones dos Santos Neves Institute (IJSN) through visual photo-interpretation of KOMPSAT-3/3A optical satellite imagery ( $0.5\text{ m/pixel}0.5\backslash text\{\; m/pixel\}$ ) collected between 2019 and 2020.

To replace slow and manual QGIS-based data retrieval, we developed an automated custom API workflow:

1. Sampling Strategy: Coordinates were sampled from an official database of CAR rural properties across the state of Espírito Santo (ES), Brazil.
2. Geographic Stratification: Coordinates were stratified across all 10 official micro-regions defined by the Secretariat for Development of Espírito Santo (SEDES), sampling exactly 1,000 coordinates per micro-region to preserve geographic and environmental diversity.
3. Automated Querying: For each coordinate, the API queries GeoBases Web Map Services (WMS) to download centered $2048\times 20482048\; \backslash times\; 2048$ pixel image tiles (covering $1,024\text{ m}\times 1,024\text{ m}1,024\backslash text\{\; m\}\; \backslash times\; 1,024\backslash text\{\; m\}$ on the ground) paired with shapefile thematic maps rasterized into segmentation masks.

Dataset Splits (60% / 20% / 20%)

The 10,000 sample pairs are stratified across the 10 micro-regions:

• satellite_train & mask_train: 6,000 tiles (600 per micro-region)
• satellite_val & mask_val: 2,000 tiles (200 per micro-region)
• satellite_test & mask_test: 2,000 tiles (200 per micro-region)

---

Semantic Class Taxonomy

The reference maps from IJSN originally feature 15 base categories and 24 fine-grained subcategories. In consultation with IDAF environmental management specialists, we grouped these into five CAR-oriented thematic classes:

Class ID	Class Name	Description	Original GeoBases Categories	Portuguese Term
0	Vegetation Areas	Native forests, wetlands, mangroves, marshes, restinga, and high-altitude rupestrian grasslands.	2, 4, 5, 6, 7, 8	Áreas de Vegetação
1	Agropastoral Areas	Pasturelands, cultivated fields (coffee, sugarcane, papaya, banana, etc.), prepared agricultural land, and regeneration areas.	1, 3	Áreas Agropastoris
2	Infrastructure	Houses, buildings, paved roads, dirt roads, sidewalks, mineral extraction sites, and constructed surfaces.	10, 11, 12, 13, 15	Infraestrutura
3	Water Bodies	Rivers, streams, lagoons, reservoirs, and visible surface-water bodies.	14	Corpos d'Água
4	Macega	Unmanaged or transitional vegetation generally composed of dense, coarse, dry tall grasses and tangled shrubs in uncultivated areas.	9	Macega

Note on Macega: Macega was preserved as a standalone class due to its distinct ecological and institutional significance in rural land monitoring, representing land in transition that is neither consolidated native forest nor actively managed agropastoral land.

---

Dataset Structure & Features

The dataset is organized as a DatasetDict containing distinct splits for satellite tiles (satellite_*) and segmentation ground-truth masks (mask_*):

DatasetDict({
    mask_train: Dataset({ features: ['image', 'filename', 'latitude', 'longitude', 'municipio', 'microestad'], num_rows: 6000 }),
    mask_val: Dataset({ features: ['image', 'filename', 'latitude', 'longitude', 'municipio', 'microestad'], num_rows: 2000 }),
    mask_test: Dataset({ features: ['image', 'filename', 'latitude', 'longitude', 'municipio', 'microestad'], num_rows: 2000 }),
    satellite_train: Dataset({ features: ['image', 'filename', 'latitude', 'longitude', 'municipio', 'microestad'], num_rows: 6000 }),
    satellite_val: Dataset({ features: ['image', 'filename', 'latitude', 'longitude', 'municipio', 'microestad'], num_rows: 2000 }),
    satellite_test: Dataset({ features: ['image', 'filename', 'latitude', 'longitude', 'municipio', 'microestad'], num_rows: 2000 })
})

Feature Definitions

• image (PIL.Image.Image): High-resolution tile ( $2048\times 20482048\; \backslash times\; 2048$ pixels, RGBA/RGB format for satellite imagery; single-channel/RGB representation for segmentation masks).
• filename (str): Unique identifier formatted by CAR property code (e.g., ES-3200102-36D5A14EF3224FC99B38DB4E06604494.tif).
• latitude (float): SIRGAS 2000 UTM projected Y-coordinate of the tile center.
• longitude (float): SIRGAS 2000 UTM projected X-coordinate of the tile center.
• municipio (str): Municipality name in Espírito Santo (e.g., Afonso Claudio).
• microestad (str): Official SEDES micro-region name (e.g., Sudoeste Serrana).

---

Getting Started & Python Usage

Due to the dataset volume (+300 GB), you can choose between Streaming Mode (zero disk footprint) and Offline Download / Chunked Loading.

Method 1: Streaming Mode (Recommended for Limited Storage)

Streaming mode lets you iterate through satellite tiles and masks without downloading the 300 GB dataset to disk:

from datasets import load_dataset
import numpy as np

# Load in streaming mode (no full download)
ds_sat = load_dataset("laicsiifes/IntegraCAR-LULC-10K", split="satellite_train", streaming=True)
ds_mask = load_dataset("laicsiifes/IntegraCAR-LULC-10K", split="mask_train", streaming=True)

# Pair samples on the fly
paired_stream = zip(ds_sat, ds_mask)

for i, (sat_sample, mask_sample) in enumerate(paired_stream):
    assert sat_sample["filename"] == mask_sample["filename"]
    
    sat_img = sat_sample["image"].convert("RGB")
    mask_arr = np.array(mask_sample["image"])
    
    print(f"Sample #{i}: {sat_sample['filename']} | Location: {sat_sample['municipio']}")
    print(f"  Satellite shape: {sat_img.size} | Mask classes present: {np.unique(mask_arr)}")
    
    if i >= 4:
        break

---

Method 2: Offline PyTorch Dataset (For High-Performance GPU Training)

If you have downloaded the shards to local storage, you can use a PyTorch Dataset wrapper:

from datasets import load_dataset
import numpy as np
from PIL import Image
import torch
from torch.utils.data import Dataset, DataLoader

# Specify custom cache directory on high-capacity drive
CACHE_DIR = "/path/to/large_storage/huggingface_cache"

# Load specific split (e.g., train split)
ds = load_dataset(
    "laicsiifes/IntegraCAR-LULC-10K", 
    cache_dir=CACHE_DIR
)

class IntegraCAR10KDataset(Dataset):
    def __init__(self, dataset_dict, split="train", transform=None):
        self.sat_ds = dataset_dict[f"satellite_{split}"]
        self.mask_ds = dataset_dict[f"mask_{split}"]
        self.transform = transform
        
        assert len(self.sat_ds) == len(self.mask_ds), "Split sizes do not match!"

    def __len__(self):
        return len(self.sat_ds)

    def __getitem__(self, idx):
        sat_item = self.sat_ds[idx]
        mask_item = self.mask_ds[idx]
        
        # Verify alignment
        assert sat_item["filename"] == mask_item["filename"]
        
        # Extract images
        sat_img = sat_item["image"].convert("RGB")
        mask_img = np.array(mask_item["image"])
        
        metadata = {
            "filename": sat_item["filename"],
            "municipio": sat_item["municipio"],
            "microregion": sat_item["microestad"],
            "coords": (sat_item["latitude"], sat_item["longitude"])
        }
        
        if self.transform:
            sat_img, mask_img = self.transform(sat_img, mask_img)
            
        return sat_img, mask_img, metadata

# Example: DataLoader with multiple workers
train_dataset = IntegraCAR10KDataset(ds, split="train")
train_loader = DataLoader(
    train_dataset,
    batch_size=4,
    shuffle=True,
    num_workers=4,
    pin_memory=True
)

---

Experimental Baselines (SIBGRAPI 2026)

In our SIBGRAPI 2026 paper, baseline experiments were established on IntegraCAR-LULC-500, the geographically stratified 500-sample subset of this collection:

• U-Net equipped with an EfficientNet-B5 encoder ( $256\times 256256\; \backslash times\; 256$ input resolution).
• DeepLabv3 equipped with a ResNet-50 encoder ( $512\times 512512\; \backslash times\; 512$ input resolution).

Baseline Performance on IntegraCAR-LULC-500

Model	Encoder	Input Patch	Step (px)	Overlap (%)	Accuracy (%)	Macro F1 (%)	mIoU (%)	Training Time (h)
U-Net	EfficientNet-B5	${256}^{2}256^2$	32	87.5%	78.0	61.3	47.8	118.3
U-Net	EfficientNet-B5	${256}^{2}256^2$	64	75.0%	82.5	62.7	53.7	42.6
U-Net	EfficientNet-B5	${256}^{2}256^2$	128	50.0%	74.0	56.8	43.1	19.7
U-Net	EfficientNet-B5	${256}^{2}256^2$	256	0.0%	74.4	65.5	48.6	3.3
DeepLabv3	ResNet-50	${512}^{2}512^2$	64	87.5%	82.5	66.8	53.7	50.6
DeepLabv3	ResNet-50	${512}^{2}512^2$	128	75.0%	81.6	65.8	52.6	14.1
DeepLabv3	ResNet-50	${512}^{2}512^2$	256	50.0%	80.9	64.3	50.8	4.2
DeepLabv3	ResNet-50	${512}^{2}512^2$	512	0.0%	81.0	65.5	52.0	1.5

Experiments executed on NVIDIA H200 NVL GPU with PyTorch 2.x.

Per-Class Metrics (DeepLabv3, 64-px Step)

Class Name	IoU (%)	Precision (%)	Recall (%)	F1-Score (%)	Class Frequency (%)
Agropastoral Areas	80.1	86.8	91.2	88.9	54.5%
Vegetation Areas	75.4	83.1	89.1	86.0	27.7%
Water Bodies	50.6	75.9	60.3	67.2	1.2%
Infrastructure	44.1	64.6	58.2	61.2	10.6%
Macega	18.2	48.6	22.6	30.8	6.0%

---

Dataset Limitations

• Photo-Interpretation Ground Truth: Labels are based on visual photo-interpretation of optical satellite scenes. Boundary uncertainties may exist in shadowed, occluded, or complex transitional zones.
• RGB Only: Imagery is restricted to RGB bands ( $0.5\text{ m}0.5\backslash text\{\; m\}$ spatial resolution). Incorporating Near-Infrared (NIR) or multi-temporal satellite series (e.g., CBERS-4A / Sentinel-2) is recommended for fine-grained discrimination of water bodies and vegetation dynamics.
• Temporal Window: Mappings correspond to the 2019–2020 temporal window provided by GeoBases.

---

Related Repositories

• Lightweight Benchmark: laicsiifes/IntegraCAR-LULC-500 (500 samples, ~10–15 GB, ideal for quick experimentation).

---

Citation Information

If you use IntegraCAR-LULC-10K, IntegraCAR-LULC-500, or the automated acquisition pipeline in your research, please cite our SIBGRAPI 2026 paper:

@inproceedings{integracar_lulc_2026,
  title     = {IntegraCAR-LULC-10K: A High-Resolution Optical Satellite Dataset for LULC Segmentation in the Brazilian Rural Environmental Registry},
  author    = {Albertino, Calebe and Lima, Gabriel M. B. and Oliveira, Vin{\'{i}}cius R. and Ribeiro, Arthur R. V. and Oliveira, Eliza K. S. and Souza, Eduardo H. P. and Dalvi, Otavio G. and Komati, Karin S. and Andrade, Jefferson O. and Boldt, Francisco A. and Paix{\~{a}}o, Thiago M.},
  booktitle = {Proceedings of the 39th SIBGRAPI Conference on Graphics, Patterns and Images (SIBGRAPI)},
  year      = {2026}
}

---

Acknowledgments

This research is supported by:

• Instituto Federal do Espírito Santo (IFES) – Serra, ES, Brazil.
• Instituto de Defesa Agropecuária e Florestal do Espírito Santo (IDAF) – Vitória, ES, Brazil.
• FAPES (Grant No. 1048/2025, Project DI 016/2025 - IntegraCAR: Integração do Cadastro Ambiental Rural no Estado do Espírito Santo; Grant 1023/2022, 055/2026, 374/2026).
• CNPq (DT-2 Grant 302726/2023-3 and Grant 407742/2022-0).