Optical and infrared remote sensing, remote sensing imagery, and target perception; includes AI analysis of optical remote sensing data, but excludes radar remote sensing unrelated to optics.
A paper in Light: Science & Applications combines dual-comb spectroscopy with computational ghost imaging and uses deep-learning reconstruction to achieve hyperspectral imaging.
Xie Peng's team at the Shanghai Institute of Optics and Fine Mechanics (SIOM) has made progress in multidimensional multiplexed optical computing, improving the parallelism of optical computing.
Why we recommend it: The original article provides specific data and details that help assess actual progress in this direction
A team from the Optoelectronics Center at the Anhui Institute of Optics and Fine Mechanics (AIOFM) published two papers in Remote Sensing, proposing the lightweight keypoint localization framework GeoAdapt and the frequency-domain low-rank adaptation segmentation method Freq-LoRA for aerial target sensing in ground-based optical remote sensing. GeoAdapt has only 5.95M parameters and achieves PCK@0.05D of 89.3%; Freq-LoRA requires just 559K trainable parameters and reaches a mean intersection-over-union of 0.904 across five simulated weather datasets, or 96.1% of full fine-tuning performance.
The Optical Remote Sensing Center of the Anhui Institute of Optics and Fine Mechanics (AIOFM) published a paper in IEEE TGRS proposing an attention-enhanced Kolmogorov-Arnold Network (AKAN) method to retrieve aerosol optical depth from polarization observations by the POSP instrument on the GF-5B satellite. The team built more than 240,000 matched samples using AERONET ground-based data. The model achieved a coefficient of determination R² of 0.9336, a root-mean-square error of 0.0492, and 94.25% of results fell within the expected error range. SHAP analysis showed that the sensitive bands and scattering-angle features identified by the model are broadly consistent with radiative transfer principles.
Guangzai Technology, a company incubated by HIOM, collaborated with the University of Cambridge, Southeast University and others to publish a paper in Sustainable Cities and Society proposing SCE-Model, a spatiotemporal carbon emission model optimized with machine learning and validated by LiDAR. Model outputs showed strong agreement with coherent differential absorption LiDAR observations (R² = 0.79; ±5 ppm), enabling hourly, block-level carbon emission mapping.