We have built an over-the-sea channel model and demonstrated D-band transmission of Quadrature Phase Shift Keying (QPSK) signals at 9 Gbaud over a 30.2 km ultra-long-distance wireless link, including a partly over-the-sea transmission channel at 128 GHz utilizing photonics-aided technology. To address nonlinear issues, we propose a quadratic convolutional neural network (QuadConvNet) in the wireless receiver to mitigate the nonlinear degradation. This approach demonstrates enhanced nonlinearity and superior learning capabilities for feature extraction, as it optimally utilizes the intrinsic high-order advantages of quadratic neurons for cognition and computation performance. It achieves a Bit Error Rate (BER) for 7 Gbaud QPSK below the 7% hard-decision forward error correction (HD-FEC) threshold of 3.8×10-3 and the 25% soft-decision forward error correction (SD-FEC) threshold of 4.2×10-2 at 9 Gbaud.