Dataset Description

Bitemporal Sentinel-1 & Sentinel-2 Dataset for Deforestation and Forestation Monitoring in Sumatra and Kalimantan

Abstract & Overview

This dataset provides a large-scale collection of 45,770 remote sensing image tiles specifically curated for bitemporal change detection and environmental conservation monitoring across the tropical regions of Sumatra and Kalimantan, Indonesia. It pairs multi-sensor optical and synthetic aperture radar (SAR) feature data with highly accurate ground truth masks to track land-cover changes. The dataset is optimized to support the training of advanced deep learning architectures, such as Vision Transformers (ViTs) and Convolutional Neural Networks (CNNs), for mapping deforestation and reforestation events.

1. Study Area

The Region of Interest (ROI) encompasses the major Indonesian islands of Sumatra and Kalimantan. These regions were selected due to their high ecological value and vulnerability to rapid land-cover transitions. The dataset covers provinces including, but not limited to, Aceh, Sumatera Utara, Riau, Jambi, Lampung, and all provinces across Kalimantan.

2. Ground Truth Generation & Methodology

The bitemporal change ground truth masks were produced via Google Earth Engine (GEE), leveraging the Dynamic World V1 (GOOGLE/DYNAMICWORLD/V1) dataset. The classification categorizes pixels into three primary temporal states based on a 70% tree-cover probability threshold (trees > 0.7).

The temporal epochs are defined as follows:

• Baseline Period (t1): Median tree cover between 2018-01-01 and 2021-01-01.

• Post-Event Period (t2): Median tree cover between 2022-01-01 and 2025-01-01.

Label Encoding:

• 0 (No Change): Areas where the tree cover status remained stable between the two temporal epochs.

• 1 (Deforestation): Areas classified as trees in t1 but transitioned to non-trees in t2 (Tree Loss).

• 2 (Forestation): Areas classified as non-trees in t1 and transitioned to trees in t2 (Tree Gain).

• -1 (NoData): Missing or masked pixels in either observation epoch.

3. Feature Space (Input Imagery)

The input imagery consists of pre-processed, cloud-masked satellite composite patches capturing both Sentinel-2 (Optical) and Sentinel-1 (SAR) data, alongside algorithmically derived spectral indices. The optical composite is generated using median values of the COPERNICUS/S2_SR_HARMONIZED collection, filtered for a cloudy pixel percentage of less than 20%, and processed with an algorithmic cloud mask using the Scene Classification Layer (SCL). The SAR composite utilizes the median of COPERNICUS/S1_GRD Interferometric Wide (IW) swath data.

Each image tile contains 10 feature bands stacked and exported in float32 format. Missing data or masked pixels are filled with a default NoData value of -1. The 10 bands are ordered and derived as follows:

• Spectral Bands (Sentinel-2 SR):

  • Red (B4)

  • Green (B3)

  • Blue (B2)

  • NIR (B8)

  • SWIR1 (B11)

  • SWIR2 (B12)

• Derived Spectral Indices (Sentinel-2):

  • EVI (Enhanced Vegetation Index): Optimized to enhance the vegetation signal with improved sensitivity in high biomass regions.

  • NBR (Normalized Burn Ratio): Calculated utilizing NIR (B8) and SWIR2 (B12) to detect burn severity and structural changes in forest cover.

• SAR Backscatter (Sentinel-1):

  • SAR (VH Backscatter): The Cross-polarized (VH) backscatter coefficient, which is highly sensitive to vegetation volume scattering and canopy structure.

• Derived SAR Index (Sentinel-1):

  • RVI (Radar Vegetation Index): Derived using both VV and VH polarizations to measure vegetation growth dynamics independent of cloud cover.

4. Dataset Characteristics and Metadata

To maintain spatial consistency, all image tiles are cropped to exactly 256x256 pixels. The geographic coordinate reference system is EPSG:4326, representing a spatial resolution of approximately 10 meters per pixel (or 0.0000898 degrees).

To assist with data loading, spatial filtering, and exploratory data analysis, two companion CSV files are included in this repository:

• general_info.csv: Contains static metadata, spatial resolution parameters, band names, and aggregate statistical metrics (min, max, mean, standard deviation) across the entire dataset.

• per_file_coordinates.csv: Contains the unique spatial attributes for every individual file, providing exact bounding box parameters and centroid coordinates. This is highly useful for mapping tiles back to their absolute geographic location or performing localized filtering.

5. Intended Use and Applications

This dataset is designed to benchmark machine learning algorithms for remote sensing. The 256x256 tile size is natively optimized for ingestion into modern semantic segmentation and change detection algorithms, facilitating robust and automated monitoring of tropical forest dynamics.

6. Code & Reproducibility (Google Earth Engine Scripts)

To ensure full scientific transparency and reproducibility, the Google Earth Engine scripts used for spatial processing, composite creation, and ground truth mask generation are publicly accessible via the following links:

• Baseline Feature Composite Script (t1: 2018–2020): GEE Script t1

• Post-Event Feature Composite Script (t2: 2022–2024): GEE Script t2

• Bitemporal Change Mask Script (Ground Truth): GEE Script Mask