Source-aware reinforcement-learning optical-channel selection for compact VIS-NIR sensing with botanical validation
Mingkun Zhang, Chao Ma, Yuxiang Li, Jiayu Huang, Mingtong Du, Huawei Niu, Jianwei Ma · Optics Express · 2026
Compact visible-near-infrared (VIS-NIR) sensors require a few optical passbands that remain informative under source and acquisition variation, but conventional wavelength selectors usually optimize classifier accuracy after dense spectra have already been collected. We introduce a source-aware reinforcement-learning (RL) optical-channel selector that treats wavelength choice as an early-stage sensor-design problem. The key innovation is a policy-gradient reward derived from a source-robust optical utility: blocked-source macro-F1, balanced accuracy, Fisher separability, source consistency, inter-channel redundancy, and channel cost are optimized in one objective, while a no-repeat action mask, entropy regularization, and weighted consensus aggregation convert stochastic selections into stable passband proposals. A 31-channel Pseudostellariae Radix-Codonopsis Radix VIS-NIR dataset covering nominal 360-980 nm measurements is used only as a botanical validation case and is evaluated with leave-one-source-out testing over four source blocks. With three channels, the proposed selector achieved a macro-F1 of 0.9580, compared with 0.9766 for the full 31-channel reference. With five channels, it achieved performance comparable to CARS, while CARS and SPA remained competitive or performed better under some other channel budgets. A 15-channel consensus policy reached a macro-F1 of 0.9733. These results position the proposed method as a source-aware compact sensing policy rather than a universally superior optical-channel selector. The final output is not a classifier-specific feature list, but a compact optical-channel specification with stable candidate centers at 455, 465, 730, 750, and 770 nm for simplified VIS-NIR sensing.