Background <p>To enable rapid radionuclide identification in nuclear emergency scenarios, we propose a lightweight residual convolutional neural network (CNN) tailored for portable γ-ray spectrometers.</p> Methods <p>A high-fidelity Geant4 model of a LaBr₃:Ce detector is constructed, and γ-ray energy spectra are simulated for six single nuclides (²²Na, ⁶⁰Co, ¹³³Ba, ¹³⁷Cs, ¹⁵²Eu, ²²⁶Ra), which are further extended to dual- and triple-nuclide mixtures to match typical nuclear emergency conditions. One-dimensional spectra are reshaped into 64 × 64 matrices through serpentine mapping and then fed into a lightweight residual CNN with an embedded channel attention mechanism to enhance the extraction of key spectral features.</p> Results <p>The proposed model achieves perfect performance for single-nuclide identification (Precision = Recall = F1 = 100.00%), and F1-scores of 99.59% and 98.70% for dual- and triple-nuclide mixtures, respectively. With only approximately 3.4 × 10⁵ trainable parameters, the network maintains high accuracy under low-count rates and severe peak overlap while remaining suitable for real-time deployment on edge devices.</p> Conclusion <p>These results demonstrate a practical and efficient solution for intelligent γ-ray spectrum-based nuclide identification, providing technical support for portable γ-ray spectrometers in nuclear emergency monitoring.</p>