Earth observation is the business of turning photons reflected or emitted by the planet into decisions, and the current research frontier sits at the two ends of that chain: infrared and single-photon detectors that see bands silicon cannot, and onboard neural processing that compresses or classifies data before it ever reaches a downlink.
Summary
Earth observation (EO) covers optical, hyperspectral and radar imaging of the planet from orbit or aircraft, plus the processing chain that converts raw measurements into usable products such as land cover maps, crop traits or vessel detections. Physically it is a photon budget problem. A sensor collects a limited number of photons per pixel per revisit, the detector adds noise, the atmosphere and platform motion blur the result, and a finite downlink then has to carry whatever survives. Every meaningful improvement is an attack on one of those terms.
The source base here clusters into three layers. The front end is detector and optics physics: mid-infrared upconversion detection reaching a noise equivalent power of 0.3 fW/Hz^1/2, described as at least a tenfold improvement over prior results; superconducting nanowire single-photon detectors made from tungsten germanide with saturated internal detection efficiency out to 29 micrometres; and single-pixel or ghost-imaging schemes that trade spatial detector arrays for computation, including mid-infrared computational ghost spectroscopy at 0.62 cm^-1 resolution and self-supervised reconstruction for near-infrared single-photon single-pixel imaging. A second layer is measurement-theoretic: optically pre-processing light before detection to beat shot-noise limits on high spatial frequencies, demonstrated with a fivefold Fisher information gain in microscopy, and quantum-resource hierarchies for LiDAR range and velocity estimation under structured noise.
The third layer is where EO actually operates today: machine learning on downlinked or onboard data. Detectors tuned to the statistics of satellite imagery, such as a ship detector that shifts its feature pyramid from strides 8/16/32 to 4/8/16 to satisfy a Nyquist-derived representability condition for narrow vessels; lightweight classifiers sized for satellite compute, reaching over 94 per cent on EuroSAT and 98 per cent on PatternNet at roughly 852 mW on neuromorphic hardware; diffusion-based synthesis of training data for newly launched SAR sensors that lack labelled archives; and neural codecs designed for machine perception and non-standard modalities such as hyperspectral cubes under bandwidth and power budgets.
The parameters that decide EO outcomes are therefore: photons per pixel and detector noise in the band of interest, cooling requirement (a hard platform constraint for superconducting detectors), downlink bits per pass, watts available for onboard inference, and the availability of labelled data matched to each new sensor’s resolution, polarisation and frequency band.
Evidence base
- Mid-infrared upconversion detection with a 36-fold cavity pump enhancement reached 22 per cent peak conversion efficiency and a noise equivalent power of 0.3 fW/Hz^1/2, at least tenfold better than prior results, at room temperature (29 May 2026).
- Tungsten germanide SNSPDs showed saturated internal detection efficiency at wavelengths up to 29 micrometres with a material system compatible with large-scale fabrication (17 June 2026).
- Existing infrared single-photon detectors require 0.08 to 0.9 K and advanced cryogenics; a ferromagnet/superconductor bilayer is proposed to raise this to 3.75 K (17 June 2026).
- ResNet-GLUSE exceeded 94 per cent accuracy on EuroSAT and 98 per cent on PatternNet with 33x fewer parameters and 27x fewer FLOPs than MobileViT (99 per cent), running at 852.30 mW on Akida neuromorphic hardware (23 July 2026).
- Newly launched SAR satellites cannot reuse existing models because of differing resolution, polarisation and frequency band, motivating LoRA-tuned diffusion generation with attention distillation of sensor-specific speckle and texture (30 June 2026).
- Quantum LiDAR advantage is conditional on noise structure: separable squeezing wins under loss and thermal background, twin beams only under correlated jamming with an adaptively optimised receiver (17 June 2026).
Open questions
- Does the fivefold high-spatial-frequency Fisher information gain from Fourier Domain Division survive atmospheric turbulence and platform motion at orbital range, or is it confined to microscopy geometry?
- Can the proposed ferromagnet/superconductor 3.75 K operating point be fabricated and, if so, does that temperature bring mid- and long-wave infrared single-photon detection within a realistic satellite thermal and mass budget?
- Do models trained on diffusion-generated cross-sensor SAR data match models trained on real labelled data from the target sensor, and by how much do they degrade on operational rather than benchmark scenes?
- Is the accuracy given up by lightweight onboard classifiers (94 per cent EuroSAT versus 99 per cent for MobileViT) acceptable for tasking and triage decisions, or does it force downlink of the data the onboard model was meant to filter?
Assessment drafted 2026-08-31 from up to 18 KB sources using the technology-scorecard framework; scores are a draft read pending review.