In this paper, the basic principle and reconstruction method of random filter spectral coding-decoding are discussed. According to the automatic feature extraction mechanism of a deep learning undercomplete autoencoder, a pixel mapping variable-resolution spectral imaging reconstruction network with high reconstruction accuracy and low delay is constructed. The parallel training of a 2×2 and 4×4 pixel array spectral reconstruction network is implemented by transforming the pixel mapping relationship. Finally, the network’s performance is verified by the remote sensing data with 512×616 with 120 bands spectral images. For a 2×2 pixel array with 40 bands, the reconstruction PSNR is 53 dB, the reconstruction MSE is less than 0.002, and the reconstruction time is 0.85 s. For a 4×4 pixel array with 120 bands, the reconstruction PSNR is 64 dB, the reconstruction MSE is less than 10^-5, and the reconstruction time is 0.5 s. The experimental results show that the pixel mapping variable-resolution spectral imaging reconstruction network has the dynamic transformation performance of high accuracy and low delay.