This repository contains the initial Python code used in the image processing pipeline. The aim is to extract a vectorised coastline from satellite imagery, enabling automated monitoring of coastlines for erosion. Once validated, the pipeline will be ported to C++ and deployed onboard the satellite.
The image processing workflow is designed to handle real satellite data and convert it into a usable, high-resolution coastline vector. The pipeline must be fast, lightweight, and capable of processing images onboard with minimal error, along complex or cloud-obstructed coastlines.
The project is still under active development, but consists of the following major components:
Identifies and masks out clouds from imagery to ensure downstream processing only uses valid land/water data.
This technique aims to identify and discard regions that are either entirely land or entirely water, identified by their low or high local entropy. The assumption is that coastlines exist in areas with moderate entropy, due to the mixed nature of land and water boundaries. Retaining only these areas is expected to reduce the total number of pixels that need full processing, lowering computation cost.
- Multi-scale entropy analysis: Compute entropy at varying spatial resolutions to capture coastline transitions at different scales.
- Directional entropy: Evaluate entropy along specific axes (e.g. horizontal, vertical) to better detect structured shoreline boundaries.
- Entropy-gradient fusion: Combine entropy maps with image gradients (e.g. Sobel edge detection) to more accurately isolate transition zones.
- Spectral-band entropy: Perform entropy analysis on spectral indices (e.g. NIR, NDWI) to exploit clearer land-water separation.
- Lightweight ML-based filtering: Use a simple classifier trained on entropy, gradient, and spectral features to discard low-value regions efficiently.
The goal is to train a UNET model to directly extract coastlines, replacing manual or spectral segmentation. However, current performance is limited. If this approach proves unreliable, the fallback use case is to use UNET for general land-water segmentation, supporting downstream vectorisation through more traditional means.
Applies spectral and clustering-based techniques to efficiently and accurately segment the image and extract the coastline.
Approaches under evaluation:
- NIR Exploitation: Water absorbs NIR strongly and appears black, whilst land does not. This spectral difference is used to segment land and water, either directly or via indices like NDWI.
- K-Means Clustering: Groups pixels based on colour or spectral similarity to separate land and water.
- Contrast Stretching: Enhances the dynamic range of input imagery, particularly useful in hazy or low-contrast scenes.
- Binarisation: Converts segmented output to a binary land/water mask suitable for vector extraction.
Once a clean binary coastline mask is available, the Marching Squares algorithm is used to extract a sub-pixel-accurate vector of the coastline. This improves precision in coastline tracking and reduces the appearance of jagged or aliased boundaries.
After coastline extraction, further processing and analysis can be applied:
- Noise Removal: Exclude small or irrelevant features (e.g. lakes, rivers, artefacts) using filters based on blob size, location, or external map overlays.
- Smoothing: Clean up jagged vector edges using curve fitting or line simplification algorithms.
- Temporal Differencing (Planned): Compare coastline vectors over time to track erosion, deposition, or other long-term changes.
- Erosion Quantification: Calculate displacement metrics, retreat rates, and identify areas of accelerated coastal change.
- Developing and validating the alternate coastline extraction pipeline
- Evaluating and benchmarking UNET versus entropy-based segmentation
- Consolidating code modules into a single pipeline for integration
- Assembling a representative dataset consistent with the onboard satellite imager for validation and stress-testing
- UNET performance is currently insufficient for reliable coastline vector extraction. It may be replaced or augmented by alternative techniques if accuracy cannot be improved.
- Dataset limitations are hindering tuning and robustness checks. Access to realistic satellite data is needed to test segmentation and thresholding methods under diverse conditions.
- Entropy segmentation remains unvalidated and may not generalise well across all geographies or lighting conditions.
- Cloud masking implemented
- Marching Squares implemented
- Alternate coastline extraction in progress
- UNET and entropy segmentation under evaluation
- Pipeline integration and modular cleanup
- Testing with representative satellite imagery
- Final C++ port (to be developed in the OirthirSAT C++ repository)
Developed by the OirthirSAT Software Team.
To be determined.