Winter bamboo shoot detection is often confined to two challenges in conventional single-domain feature analysis at present: 1) Electromagnetic wave reflection signals emitted and received by detectors exhibit unclear and indistinct features, thus impacting the signal analysis. 2) It is extremely difficult to accurately distinguish echo signals from winter bamboo shoots and bamboo rhizomes, due to their similar physical properties and structural traits. In this study, a RES-ICSE-CBAM network architecture was constructed with strong adaptability. A reliable multi-domain feature fusion strategy was integrated for signal processing. Effective signals were collected from the time frequency, autocorrelation and wavelet domain for the feature diversity. Specifically, ICSE module of feature enhancement achieved in the accurate and efficient identification between underground winter bamboo shoots and bamboo rhizomes in complex soil environments. Strong robustness was compensated for the loss of key detailed information during feature extraction, leading to the overall performance of feature representation and signal differentiation. Residual network (RES) was introduced to alleviate the common gradient vanishing in the process of network training. An improved convolutional block attention module (CBAM) was embedded into the network structure to optimize feature extraction. The extraction of key signal features was enhanced for feature selection accuracy using dual complementary, including channel and spatial attention mechanism. Meanwhile, three signal processing approaches were adopted to conduct multi-dimensional analysis of original time-domain signals, such as short-time fourier transform (STFT), autocorrelation calculation (AC) and wavelet transform (WT). Among them, STFT was utilized to convert original time- into time-frequency domain signals. Time-frequency information was obtained, as signal frequencies varied dynamically over time. AC was employed to explore the autocorrelation for potential change of the signals. WT was applied to perform multi-scale decomposition of the signals with high precision. Detailed feature information was extracted to distinguish subtle signal differences at various scales. Stacking multi-domain fusion was combined to integrate valuable complementary information from the three feature domains. Each domain was fully utilized to avoid information waste. As such, subtle and weak signals were identified in complex environments. Experimental results further demonstrated that the overall recognition accuracy of multi-domain feature fusion reached 84.44% in practical tests, indicating better stable performance. Specifically, the recognition accuracies were 86.67%, 83.75% and 82.92%, respectively, for soil, winter bamboo shoot and bamboo rhizome signals. The multi-domain feature fusion achieved higher recognition accuracy than conventional single-domain feature analysis. The favorable reliability of the model was achieved to identify winter bamboo shoots and bamboo rhizomes.