Spectrum sensing is an indispensable core part of cognitive radio dynamic spectrum access (DSA) and a key approach to alleviating spectrum scarcity in the Internet of Things (IoT). The key issue in practical IoT networks is robust sensing under the coexistence of low signal-to-noise ratios (SNRs) and non-Gaussian impulsive noise, where observations may be distorted differently across feature modalities, making conventional fusion unstable and degrading detection reliability. To address this challenge, the generalized Gaussian distribution (GGD) is adopted as the noise model, and a multimodal fusion framework termed BCAM-Net (bidirectional cross-attention multimodal network) is proposed. BCAM-Net adopts a parallel dual-branch architecture: a time-frequency branch that leverages the continuous wavelet transform (CWT) to extract time-frequency representations, and a temporal branch that learns long-range dependencies from raw signals. BCAM-Net utilizes a bidirectional cross-attention mechanism to achieve deep alignment and mutual calibration of temporal and time-frequency features, generating a fused representation that is highly robust to complex noise. Simulation results show that, under GGD noise with shape parameter $\beta =0.5$, BCAM-Net achieves high detection probabilities in the low-SNR regime and outperforms representative baselines. At a false alarm probability $P_f=0.1$ and SNR of $-14$ dB, it attains a detection probability of $0.9020$, exceeding the CNN-Transformer, WT-ResNet, TFCFN, and conventional CNN benchmarks by 5.75%, 6.98%, 33.3%, and 21.1%, respectively. These results indicate that BCAM-Net can effectively improve spectrum sensing performance in low-SNR impulsive-noise scenarios, and provides a lightweight, high-performance solution for practical cognitive radio spectrum sensing.