ra-hk1/spillvision-deeplab-keras

SpillVision DeepLabV3+ Keras Models

This repository contains Keras-based DeepLabV3+ model checkpoints developed for oil spill segmentation from satellite imagery as part of the SpillVision project.

Model Purpose

The models are used for semantic segmentation. The goal is to identify oil spill regions in satellite images and generate pixel-level masks.

The predicted mask can be used to calculate spill coverage, spill area, centroid location, risk level, and environmental response information.

Repository Files

File	Description
best_deeplab.keras	Main trained DeepLabV3+ Keras model checkpoint
best_deeplab_stage1.keras	Stage 1 training checkpoint
best_deeplab_finetuned.keras	Fine-tuned DeepLabV3+ checkpoint

Intended Use

This model is intended for oil spill segmentation, remote sensing research, environmental monitoring prototypes, academic demonstrations, and integration with geospatial risk analysis pipelines.

Not Intended For

This model should not be used as the only source of truth for emergency response decisions.

For real-world use, predictions should be reviewed by domain experts and validated with additional data sources such as field reports, satellite imagery, weather data, ocean currents, and official environmental monitoring systems.

Model Architecture

The model is based on a DeepLabV3+ semantic segmentation architecture implemented with Keras and TensorFlow.

Input

Expected input is a satellite image prepared using the same preprocessing steps used during training.

Typical preprocessing may include resizing the image and normalizing pixel values.

Output

The model outputs a segmentation mask where predicted pixels represent possible oil spill regions.

Typical post-processing may include thresholding, binary mask generation, contour extraction, coverage calculation, and geospatial risk analysis.

Project Pipeline

In the full SpillVision system, the model is connected to a larger workflow:

1. Upload or select a satellite image
2. Run DeepLabV3+ segmentation
3. Generate the predicted oil spill mask
4. Calculate spill metrics
5. Estimate risk level
6. Store results in the database
7. Visualize cases on an interactive map
8. Generate AI-supported environmental reports

Limitations

The model may be affected by image resolution, cloud cover, sensor differences, missing georeferencing information, small spill regions, and domain shift between training data and new satellite images.

Predictions should always be interpreted carefully.

Ethical and Environmental Considerations

This model is designed to support environmental monitoring and early detection of possible oil spills.

Automated predictions can contain errors. Human review and official validation are required before making environmental, legal, or emergency response decisions.

Citation

SpillVision: AI-powered oil spill detection and risk analysis from satellite imagery.

Author

Developed by Rana Kenani as part of the SpillVision oil spill detection and monitoring project.