NASA and commercial satellite imagery provide a continuous, ground-truth view of physical energy infrastructure — crude storage tanks, refinery throughput, pipeline corridors, LNG terminals, and power generation assets. This layer captures changes in the real world before they register in official reports or exchange data, giving the system an informational edge grounded in observable reality.
Multispectral and synthetic aperture radar (SAR) imagery is processed on AWS to extract structured signals from raw visual data. Shadow analysis of floating-roof crude tanks estimates inventory fill levels. Thermal anomalies flag refinery startups and shutdowns. Vessel positioning at marine terminals surfaces import and export flow ahead of official customs data.
Floating-roof tanks cast measurable shadows that vary with fill level. Satellite revisit cycles of 1–3 days allow near-continuous inventory tracking at Cushing, OK and Gulf Coast terminals — producing independent storage estimates ahead of the EIA weekly report and surfacing inventory builds or draws before they hit the tape.
Thermal infrared imagery detects heat signatures from distillation columns, fluid catalytic crackers, and flare stacks. An unplanned shutdown registers as a rapid thermal cool-down across the facility footprint — a signal that feeds directly into heating oil and gasoline crack spread modelling, typically 48–72 hours before PADD-level supply data is published.
AIS vessel data combined with satellite imagery of LNG berth occupancy tracks export loading cycles at Sabine Pass, Corpus Christi, and Freeport. Berth congestion or unexpected vessel diversions signal supply disruptions or demand shifts in European and Asian LNG markets — feeding the natural gas modelling layer with forward-looking export flow data.
redr labs builds deep learning architectures purpose-built for the non-linear, non-stationary behaviour of energy markets and the physical data that feeds them. Recurrent and transformer-based sequence models learn temporal dependencies across satellite revisit cycles, storage draws, and price action that classical linear regressions and rule-based heuristics consistently miss.
Convolutional networks are trained directly on multispectral and SAR imagery to extract features — tank shadow geometry, thermal gradients, deformation fringes — without hand-engineered proxies. These learned representations feed downstream sequence models, creating an end-to-end pipeline from raw pixel data to trend and anomaly signal.
Our research stack is built on PyTorch, with experiment tracking and distributed training across GPU clusters. The goal is a class of models that can surface correlation drift, regime transitions, and structural breaks in energy fundamentals — patterns that are real but too non-linear for classical statistical methods to reliably capture.
LSTM and attention-based sequence models trained on multi-year energy time-series capture long-range dependencies and regime shifts across natural gas, crude, and power markets that classical autoregressive models cannot represent.
CNNs trained on multispectral, thermal, and SAR imagery learn to detect tank fill state, refinery heat signatures, and terrain deformation directly from pixel data — replacing brittle hand-tuned image-processing rules with learned representations.
Model development runs on PyTorch with distributed GPU training, enabling rapid iteration on large imagery and time-series datasets and supporting experimentation with deeper architectures than a single-node workflow would allow.
Autoencoder reconstruction error and non-linear dimensionality reduction surface anomalies and structural breaks in cross-commodity correlations — an orthogonal signal source to classical statistical change-point detection.
satellite data science
Interferometric synthetic aperture radar (InSAR) measures ground deformation at millimetre precision by comparing the phase of successive radar passes over the same terrain. Applied to pipeline corridors, storage sites, and well pads, this surfaces subsidence, uplift, and slope movement that signal subsurface pressure changes, leaks, or structural risk long before it is visible on the surface.
Infrared channels extend this picture into the thermal domain. Long-wave and shortwave infrared bands are radiometrically calibrated and corrected for atmospheric absorption and surface emissivity to convert raw sensor counts into physically meaningful temperatures — the basis for detecting flare activity, combustion anomalies, and equipment-level heat signatures across a facility footprint.
Time-series InSAR stacking across repeat radar passes builds a continuous deformation history for pipeline right-of-way, injection wells, and storage cavern sites. Anomalous subsidence or uplift rates flag integrity risk, ground instability, or reservoir pressure changes — a physical monitoring layer independent of self-reported operator data.
Raw thermal sensor counts are converted to at-surface brightness temperature through radiometric calibration, then corrected for atmospheric water vapour absorption and target emissivity. This physically grounded temperature field is what makes flare detection, combustion efficiency estimates, and equipment-level thermal anomalies quantitatively reliable rather than merely visual.
Raw scenes are orthorectified, cloud-masked, and pan-sharpened before spectral indices are computed across visible, near-infrared, and shortwave bands. This processed imagery becomes the shared input layer for both classical feature extraction and the convolutional models trained to detect infrastructure state changes directly from pixels.