LizardNet: Lizard Connectome-Inspired Network for Fast Few-Shot Hyperspectral Classification

Abstract

Hyperspectral image (HSI) classification requires spectral selection, boundary-sensitive spatial processing, and low model complexity. We present LizardNet, a lightweight neuro-dynamic framework whose architecture draws on three organizational features of the reptilian visual pathway: retinal oil-droplet filtering, horizontal-cell lateral inhibition, and shallow tectal integration. LizardNet implements these features with three modules. The connectomics-inspired sparse spectral gating (SSG) module uses an ODE-inspired steady-state score and Top- $K$ hard gating. The lateral inhibition and anisotropic diffusion (LIAD) module applies a partial differential equation (PDE)-based, boundary-aware spatial update. The tectal low-rank multipath integration (TMI) module uses nuclear-norm regularization to constrain multipath fusion. On Indian Pines (IP) with 5% training samples, LizardNet obtains 94.63% overall accuracy (OA); on Trento, it obtains 98.75% OA. CPU-only FP32 measurements on a Jetson Orin Nano 8 GB give 25.14 samples/s, 39.78-ms mean latency, and 0.191-J incremental energy per sample. Under the same protocol, DGPF-RENet gives 3.18 samples/s, 314.22-ms mean latency, and 1.854-J incremental energy per sample. These results quantify the accuracy and computational cost of using explicit spectral, spatial, and low-rank constraints for HSI classification. The code and trained models are available at: https://github.com/yuhua666/LizardNet