Blog: Beyond RGB: What Satellite Imagery Really Contains
Figure: Conceptual illustration of satellite image dimensions (AI-generated based on author description)
When we look at a photograph, we see colour, shape, and shadow captured in three channels: red, green, and blue. A satellite image looks similar on the surface, but it is a fundamentally different kind of data. Understanding that difference is the first step toward analysing it properly.
A satellite image has three key dimensions. The first is spatial resolution. Every pixel in a satellite image corresponds to a precise latitude and longitude on the ground. This means a pixel is not just a visual element, it has a geographic connotation. You can locate it on Earth, associate it with a real place, and study the spatial patterns of features across an area. The second dimension is spectral resolution. Satellites carry multiple sensors, each measuring a different part of the electromagnetic spectrum. The Sentinel-2 satellite, for example, captures 12 spectral bands. Each band records a distinct physical measurement of the Earth's surface. It doesn't just capture colour as the human eye perceives it, it captures reflectance values that correspond to real physical phenomena. The third dimension is temporal resolution. Satellites pass repeatedly over the same location, building up a historical archive. This makes it possible to study how a place changes over time, rather than just what it looks like on a specific day.
Standard computer vision (CV) approaches like convolutional neural networks have become the default toolkit in image analysis, and they handle satellite data reasonably well in some respects. They are good at extracting geometry: identifying building outlines, road networks, and land cover boundaries. They handle texture well too, and they scale across different geographies once trained. However, since they were basically designed for and trained on RGB images, they treat satellite data as though it were RGB. The reflectance values in satellite image bands are not just colour channels; they measure physical reality. The ratio of near-infrared to red reflectance, for instance, directly indicates vegetation health. A model that treats this as an arbitrary pixel value is not obeying the underlying domain logic behind the reflectance values.
Non-CV approaches have addressed this to some extent. Methods such as principal component analysis on multi-temporal difference images, or histogram-based analysis of pixel values, work directly with the spectral and temporal structure of the data rather than treating it as a picture. These approaches tend to be more physically interpretable, which means what they compute corresponds to something meaningful in the real world. But they have their own weakness: they do not generalise well. A method built for a specific sensor, task, or region often requires significant re-engineering to work elsewhere. This physical and domain grounding comes at the cost of adaptability.
The way forward is to treat a satellite image as what it actually is: a layered object containing visual context, geometric structure, and physical measurements, each of which is best handled differently. The visual and geometric components are well suited to the pattern-recognition strengths of CV models. The physical measurement component, however, need a different treatment. The proposal here is to incorporate non-trainable layers into standard CV models with fixed computational steps that apply known domain formulae to the spectral bands. These layers do not learn from data; they encode what we already know from the physical sciences. This way, the model respects the physical phenomena the sensor is measuring, rather than merely approximating them from examples.
