Contest No. 4

Remote Sensing

Title: "Urban Construction Change Detection Challenge Using High-Resolution Imagery and Multispectral Sentinel-2 Data"

Urban Construction Change Detection Challenge Using High-Resolution Imagery and Multispectral Sentinel-2 Data

 

Scope and Objective: In this challenge, a combination of high-resolution imagery and multispectral Sentinel-2 data is used to identify urban construction changes. The high-resolution (HR) images are derived from the publicly available SpaceNet-7 dataset, which is based on PlanetScope satellite imagery with an approximate spatial resolution of 4 meters, comprising multi-temporal building annotations from rapidly urbanizing areas across the globe.

In parallel, Sentinel-2 imagery offers 10-meter resolution and rich spectral coverage, including near-infrared (NIR) and short-wave infrared (SWIR) bands, which are highly valuable for distinguishing built-up surfaces from natural land covers. Combining these two data sources allows participants to analyze urban changes both spatially and spectrally, leading to more accurate identification of construction growth patterns.

 

Provided Data: The dataset includes 48 different geographic locations, each with two-time steps () approximately two years apart.

• • High-Resolution Images (HR): For each location and each time step, an RGB GeoTIFF image with a spatial resolution of approximately 4 meters is provided. These high-resolution images enable detailed analysis of urban features and precise delineation of building structures.
  • Sentinel-2 Multispectral Images: For each location and each time step, a cropped Sentinel-2 image precisely aligned with the HR imagery is included. All Sentinel-2 bands are provided in GeoTIFF format as 16-bit unsigned integer images, ensuring high radiometric fidelity for advanced spectral analysis.  The Sentinel-2 images contain 12 spectral bands:
    • Bands at 10-meter resolution: B02 (Blue), B03 (Green), B04 (Red), B08 (NIR).
    • Bands at 20-meter resolution: B05, B06, B07, B8A, B11, B12.
    • Bands at 60-meter resolution: B01, B09
  • Building Footprint Vector Files (GeoJSON): For each training location and each time step, a vector file in GeoJSON format is provided, containing building footprints extracted and annotated for each date. These vector annotations are only available for the training samples and should be used to generate binary masks representing building extents and to create change maps for model training.

(a) Sapcenet-7 RGB HR image	(b) Sentinel-2 True color image	(c) binary building mask

 

 

 

 

 

 

 

 

 

 

 

Train/Test Split: The dataset is divided into 40 geographic locations for training and 8 locations for testing. The test areas include only imagery for both time steps and do not provide any building footprint ground truth for the second time. If participants require additional validation during model development, they are encouraged to reserve a portion of the training areas as their internal validation set.

Data Access: Training and Test dataset can be access from google drive space.

 

Expected Deliverables: The required output for each test area is a binary change map in GeoTIFF format, indicating transitions from non-urban to urban (i.e., areas that have been newly built up).

Participants must submit:

• • Binary change maps (GeoTIFF format) for each test area. Each pixel should be labeled as 1 if it changed from non-urban to urban between the two-time steps, and 0 otherwise.
  • A concise report (maximum 2 pages, PDF) summarizing the model architecture, preprocessing steps, key results, and example visualizations.
  • A notebook or executable script (e.g., Jupyter Notebook, Google Colab, or Python script) that clearly demonstrates the method and can reproduce the submitted results.

All deliverables should be organized into a single ZIP file and named clearly, including the team or participant’s names.

Evaluation Metrics: Participants' submissions will be ranked based on the following criteria; this combination of metrics ensures that models are evaluated not only on their technical accuracy but also on practical efficiency and the quality of scientific contribution.

     

Metric	Description
IoU (Change Class)	Intersection over Union for the change
F1-score	The harmonic means of precision and recalls for the change class, measuring both accuracy and completeness of change detection.
Runtime	The time required to generate predictions for each test area.
Report Clarity	The scientific quality and clarity of the submitted report, including explanation of methods, visualizations, and overall presentation.

Rules and Restrictions: To ensure fairness among all participants:

• • use of any external data beyond the provided dataset is strictly prohibited.
  • Participants are allowed to submit a final submission, following the official submission schedule announced by the organizers.
  • Since this is a student competition, each team member is required to include a valid student identification card (student ID) as part of the final submission package.

Recommended Preprocessing Steps: To support participants in preparing their data effectively and ensure consistent and fair comparisons, the following preprocessing steps are strongly recommended:

1. 1. Rasterizing Building Footprints

• • Convert the provided vector building footprint files (GeoJSON) into binary raster masks.
  • These masks serve as ground truth for training.

1. 1. Generating Change Maps for Training Areas

• • Using the building masks from time steps T₁ and T₂, create binary change maps that highlight pixels transitioning from non-urban to urban (newly built-up areas).

1. 1. Data Augmentation

• • Split large images into smaller patches (e.g., 256 × 256 or 512 × 512 pixels) to increase the number of training samples and make model training more efficient.
  • Consider overlap between patches and apply common augmentations (e.g., rotations, flips, brightness adjustments) to further enrich the dataset.