AI-Powered Change Detection: How It Works

Change detection is the core engine of what we are building at ShieldNex: the part that turns raw satellite imagery into something a reviewer can act on. This post describes the intended design. ShieldNex has a research prototype, and the performance of this pipeline has not yet been benchmarked.

Multi-Temporal Analysis

Our system doesn't just look at a single image, it analyzes sequences of images captured over time. By comparing pixel-level and feature-level differences between temporal pairs, we can detect even subtle changes that would be invisible to the human eye.

Deep Learning Architecture

The approach is a convolutional neural network trained on labelled satellite image pairs. The distinction the model has to learn is between:

• Significant changes: New construction, land clearing, boundary modifications
• Natural variations: Seasonal vegetation changes, shadow differences, atmospheric effects
• Noise: Sensor artifacts, cloud cover, registration errors

Confidence Scoring

Every detected change is assigned a confidence score based on multiple factors:

1. Magnitude of spectral change
2. Spatial coherence of the change area
3. Temporal consistency across multiple observations
4. Contextual analysis of surrounding land use

From Detection to Evidence

A change above threshold is a candidate, not a finding. It goes to a human reviewer, and only an accepted candidate produces a change report, before/after imagery, change polygons, area measurements and confidence metrics, packaged with the provenance records behind them.

The two outputs people tend to overlook are the interesting ones: rejected and inconclusive. A pair of images that cloud, haze or misalignment has made unusable should return no answer rather than a confident one, and the rejections a reviewer makes are the training signal that makes the next version less wrong. We intend to keep both.