Change detection with Raster Vision
Author: Adeel Hassan
Change detection is the computer-vision equivalent of the spot-the-difference game. Given two images, the model must detect all the points at which they differ. In the context of remote sensing, these images are usually satellite or aerial images of the same geographical location at two different points in time.
Change detection has been an active research area for a long time and the literature is rich with algorithms that perform the task automatically, ranging from basic image processing techniques to present-day deep neural networks. And these algorithms have a wide range of applications including detection of deforestation, damaged buildings in the wake of disasters, and changes in urban land use. See, for example, [1] for an up-to-date review of change detection techniques.
A broad way to categorize change detection techniques is by how they model change. The direct classification approach models the change itself. In the supervised machine learning context, this requires annotations in the form of “change masks”; these masks might either be binary (change / no-change) or more specific (with classes like “forest-to-land”, “land-to-water”, etc.). The post-classification approach, on the other hand, models the underlying features (e.g. tree cover) in an image and then compares those features from both images to see what has changed. For this, one needs annotations for these features rather than for the changes. When working at the pixel-level, both approaches amount to semantic segmentation.
This blog will explore the direct classification approach to change detection using our open-source geospatial deep learning framework, Raster Vision, and the publicly available Onera Satellite Change Detection (OSCD) dataset. Raster Vision allows us to easily tackle the peculiarities of this dataset and set up a semantic segmentation training pipeline that can quickly produce a decent model.
To continue reading the full blog post please visit: https://www.azavea.com/blog/2022/04/18/change-detection-with-raster-vision/