Computer Vision for GEOINT: How ML Models Are Rewriting Satellite Imagery Analysis

Analysts used to stare at imagery for hours. A skilled photointerpreter could identify a SA-2 battery from shadow geometry and revetment spacing, but the process was slow, serial, and dependent on human stamina. The satellite constellation has grown faster than the analyst workforce. Something had to give.

Machine learning models now sit between raw imagery and the analyst's screen, handling the classification and change detection work that consumed most of the manual labor. The shift isn't about replacing photointerpreters — it's about eliminating the pixel-staring so they can focus on the part that actually requires intelligence: determining what the change means.

The pipeline architecture looks deceptively simple until you actually build it.

flowchart LR
    A[Satellite Acquisition] --> B[Preprocessing\nAtmospheric correction\nOrthorectification]
    B --> C[Change Detection\nTemporal diff\nSAR coherence]
    C --> D[Object Detection\nYOLO / DETR\nROI extraction]
    D --> E[Classification\nResNet / ViT\nObject typing]
    E --> F[Confidence Scoring\n& Triage]
    F --> G[Analyst Review\nAnomalies + context]

Preprocessing is unglamorous but load-bearing. Raw multispectral imagery contains atmospheric distortion, sensor noise, and geometric errors from orbital geometry. Atmospheric correction — converting digital numbers to surface reflectance — is non-negotiable for any cross-temporal comparison. Orthorectification removes terrain-induced displacement. Skip either step and your change detection baseline is corrupted before the model sees a single pixel.

Change detection sits at two levels. Pixel-level methods — temporal differencing, image rationing, principal component analysis across date pairs — flag regions where spectral signatures shifted. SAR coherence analysis does the same for synthetic aperture radar data, which penetrates cloud cover that makes optical imagery useless during monsoon season or over polar regions. The output isn't a classification yet; it's a set of candidate regions worth examining further.

Object detection runs on those regions. YOLO variants dominate operationally because inference speed matters when you're processing thousands of image chips per hour. DETR-based architectures offer better performance on small objects at the cost of latency — a real tradeoff when the target is a man-portable launcher in a tree line. Models trained on DOTA (Dataset for Object Detection in Aerial Images) or SpaceNet give you a foundation, but operational performance requires fine-tuning on the actual sensor, resolution, and geographic region you care about. A model trained on commercial 50cm imagery from CONUS transfers poorly to 1m SAR imagery over the Sahel.

Classification follows detection. This is where the distinctions get hard: not just "vehicle" but "wheeled APC versus tracked IFV," not just "aircraft" but "fighter versus transport based on wingspan ratio and engine pod count." Vision transformers — ViT, Swin — have pushed performance on these fine-grained tasks, particularly when you can leverage high-resolution patches. The key architectural question is whether to train a single multi-class head or a cascade of binary classifiers. Cascades are interpretable and easier to update when a new object type needs to be added without retraining everything.

Confidence scoring determines what reaches the analyst queue. The operational failure mode isn't the model missing a target — it's flooding analysts with false positives until they stop trusting the system. Calibrated uncertainty estimates, rejection thresholds tuned to acceptable false positive rates, and ensemble disagreement as a proxy for ambiguity all matter here. A model that surfaces 500 candidates with 60% average confidence creates more work than the manual process it replaced. A model that surfaces 40 candidates with 90% confidence changes the analyst's day.

The analyst review layer is where the system either earns trust or loses it. Change detection flags and classifications need to arrive with enough context — adjacent imagery, historical baseline, object metadata — that the analyst can assess the machine's reasoning rather than just accept its output. Explainability isn't academic here. When the confidence score and the analyst's intuition diverge, the analyst needs to know why the model thinks what it thinks. That's the feedback loop that makes the next training iteration better.

The technology is mature enough to deploy. The harder problems are data labeling at scale for specialized object classes, cross-domain adaptation between sensor types, and building evaluation pipelines that measure analyst throughput and not just model mAP. Optimizing for the benchmark metric without measuring whether analysts actually work faster is the most common way to ship a GEOINT ML system that nobody uses.

---

R. Tanaka researches AI systems applications in national security contexts.