Neural-based spatial-spectral sensitivity correction for pushbrum hyperspectral cameras

A neural network-based method for trainable spatial-spectral sensitivity correction of a push-broom hyperspectral sensor is proposed. Unlike traditional calibration approaches that equalize the recorded signal from a uniformly illuminated Lambertian reference target along the slit using precomputed correction coefficients, the proposed calibration is implemented as neural network layers whose parameters are jointly optimized with the classification model during training. Three types of trainable calibration layers have been developed, based on a learnable matrix of correction coefficients, its partial polynomial approximation, and vector factorization. Experimental evaluation was conducted on hyperspectral images acquired under similar but not identical capture conditions. The calibration layers were integrated into a 3D convolutional neural network and a spatial-spectral transformer. The results demonstrate a consistent improvement in classification quality compared to baseline models without calibration: the accuracy increased by 0.59% to 14.27%, and the F1-score increased by 0.17% to 10.98%, which confirms the effectiveness of the proposed layers.