CerraData-4MM - A Multimodal Dataset on Cerrado for Land Use and Land Cover Classification

CerraData-4MM is a multimodal dataset focusing on the Bico do Papagaio ecoregion within the Cerrado biome. This dataset is designed to support regional studies and applications aimed at the conservation and management of this unique and biodiverse biome, which lies in a transitional area with the Amazon forest. The dataset includes two modalities of data: multispectral patches (12 channels) and Synthetic Aperture Radar (SAR) patches (2 channels). Each modality contains 30,291 image patches of size 128x128 pixels with a spatial resolution of 10 meters. The dataset is enriched by its hierarchical organization into two levels of classes, making it a valuable resource for land use and land cover classification tasks. Read the full paper at https://doi.org/10.48550/arXiv.2502.00083.

---

About the Dataset

CerraData-4MM integrates SAR and multispectral imagery from the Sentinel-1 (S1) and Sentinel-2 (S2) satellites, respectively. The images were collected in 2022 and are aligned with reference data based on a land use and land cover map produced by the TerraClass Cerrado program in 2022. The dataset is particularly challenging at the second class level, especially regarding vegetation regeneration stages and types of agriculture, as the categories exhibit similar patterns.

Key Features:

• Two Modalities:
  • Multispectral patches (12 channels)
  • SAR patches (2 channels)
• Image Patches: 30,291 patches per modality
• Patch Size: 128x128 pixels
• Spatial Resolution: 10 meters
• Hierarchical Classes:
  • Level 1 (L1): 7 classes
  • Level 2 (L2): 14 classes

Class Hierarchy

The dataset contains two hierarchical levels of classes: Level 1 (L1) and Level 2 (L2). Below is the complete list of classes, their corresponding IDs, and their RGB colors code.

L1 Class	L1 ID	L2 Class	L2 ID	RGB
Pasture	0	Pasture (Pa)	0	206, 239, 98
Forest	1	Primary Natural Vegetation (V1)	1	22, 152, 13
Agriculture	2	Secondary Natural Vegetation (V2)	2	31, 212, 18
Mining	3	Mining (Mg)	4	176, 176, 176
Building	4	Urban area (UA)	5	223, 124, 38
Water body	5	Water body (Wt)	3	19, 50, 255
Other Uses	6	Other Built area (OB)	6	250, 128, 114
		Forestry (Ft)	7	85, 107, 47
		Perennial Agriculture (PR)	8	230, 32, 108
		Semi-perennial Agriculture (SP)	9	139, 105, 20
		Temporary agriculture of 1 cycle (T1)	10	255, 215, 0
		Temporary agriculture of 1+ cycle (T1+)	11	255, 255, 0
		Other Uses (OU)	12	117, 10, 194
		Deforestation 2022 (Df)	13	205, 0, 0

CerraData-4MM comprises sets of SAR and MSI data, followed by semantic maps for L1 and L2, respectively. The two class levels create a diverse and challenging dataset that covers categories of regeneration level, deforestation increment, as well as different types of agriculture. Each set contains 30,322 patches with a spatial resolution of 10 meters. This dataset provides a rich diversity of classes representative of the \textit{Bico do Papagaio} ecoregion. As illustrated in Figure below, L2 introduces five additional subcategories of agricultural types, one additional category for built areas, and two distinct vegetation generation classes.

---

Tutorial

First of all, you have to install the packs for this tutorial. Thus, download the file environment.yml to create a conda env.

• Installing dependencies with Conda:

conda env create -f environment.yml
conda activate cerra

• Installing dependencies with Pip:

pip install -r requirements.txt

Data Downloading

You can download the dataset using the kagglehub Python package. Below is an example of how to download the latest version of the dataset:

import kagglehub

# Download the latest version
path = kagglehub.dataset_download("cerranet/cerradata-4mm")

print("Path to dataset files:", path)

Data Loading

First, you can split the data into sub-directories for training, testing, and validating phases. Hence, consider the following code example, in which we are splitting the data into train and testing sub-directories.

import splitfolders

# Split data into Train and testing

input_folder = '../../cerradata-4mm/'
output_folder = '../../cerradata4mm_exp/'

# Ratio of split are in order of train/val/test.
splitfolders.ratio(input_folder, output_folder, seed=42, ratio=(.9, .0, .1))

Now, let's load the data! Find the function on the dataset-loader within the CerraData-4MM Experiments/util/ directory, and then import the MMDataset() to read both modalities or SARDataset() MSIDataset() to read one of the modalities. In order to exemplify how to load the data from the train subset, let's consider MMDataset() class.

import torch
import os
import sys
from torch.utils.data import DataLoader, random_split
import numpy as np
sys.path.append(os.path.abspath(os.path.join(os.path.dirname(__file__), '..', '..', '..', 'util')))
from dataset_loader import MMDataset

# GPU
## On NVIDIA architecture
device = 'cuda:0' if torch.cuda.is_available() else 'cpu'

## On Apple M chip architecture
#device = torch.device("mps")

# Path
DATA_PATH = '../../cerradata4mm_exp/train/'

# Data loading
total_data = MMDataset(dir_path=DATA_PATH, gpu=device, norm='none') # norm: "none", "0to1", "1to1"

# Split into Training and Validation phase
train_set_size = int(len(total_data) * 0.8)
val_set_size = len(total_data) - train_set_size
train_set, val_set = random_split(total_data, [train_set_size, val_set_size], generator=torch.Generator().manual_seed(random_seed))

trainloader = DataLoader(train_set, batch_size=batch_size, shuffle=True)
valloader = DataLoader(val_set, batch_size=batch_size, shuffle=False)

dataloaders = {"train": trainloader, "val": valloader}
dataset_sizes = {"train": len(trainloader), "val": len(valloader)}

How to train a model on CerraData-4MM

Let's play a bit of training U-Net on CerraData-4MM. In the following summarized example, the model is training on the first hierarchical level, i.e., 7 classes, considering both modalities, hence 14 channels.

---

Standard Benchmarks

We selected the U-Net and TransNuSeg models to experiment with the dataset. The models' performance scores are presented below. The figure presents an overview of the overall results for each baseline model across the diverse scenarios. It incorporates scores that indicate the utilization (W) and non-utilization (NW) of class weights for dataset balancing. For instance, the notation “MSI-NW-L1” denotes the modality set without class weights at the initial hierarchical level of classes. TransNuSeg outperforms U-Net in most scenarios, particularly at the L1 hierarchical level, while the ViT-based model demonstrates better multimodal learning. Combining MSI and SAR modalities improves performance compared to using either individually, but scores drop significantly for more complex L2 classes. For instance, in the Concat-NW-L1 scenario, the ViT-based model achieves an F1-score of 57.60%, whereas for the “MSI-NW-L1” and “SAR-NW-L1” scenarios, the corresponding scores are 57.45% and 38.94%, respectively. In the “MSI-NW-L2” scenario, TransNuSeg achieves an F1-score of 57.45% and an mIoU of 48.94% in L1, but these metrics decline to 53.17% and 41.90%, respectively. U-Net exhibits even more unsatisfactory performance; in the “Concat-NW-L2” scenario, its F1-score and mIoU decrease to 18.16% and 15.84%, respectively.

Read the full paper at < https://doi.org/10.48550/arXiv.2502.00083 >.

---

Citation

@misc{miranda2025cerradata4mmmultimodalbenchmarkdataset,
      title={CerraData-4MM: A multimodal benchmark dataset on Cerrado for land use and land cover classification},
      author={Mateus de Souza Miranda and Ronny Hänsch and Valdivino Alexandre de Santiago Júnior and Thales Sehn Körting and Erison Carlos dos Santos Monteiro},
      year={2025},
      eprint={2502.00083},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2502.00083},
}

---

Contact

If you have any questions, please let us know at [email protected]