Abstract: Current winter bamboo shoot detection in practical applications faces two key and prominent difficulties. First, the electromagnetic wave reflection signals emitted and received by detectors often have vague and indistinct characteristics, which affects signal analysis. Second, it is extremely hard to accurately distinguish echo signals between winter bamboo shoots and bamboo rhizomes due to their highly similar physical properties and structural features. To effectively solve these practical problems, this paper constructed a novel RES-ICSE-CBAM network architecture with strong adaptability. It also proposed a reliable multi-domain feature fusion strategy for signal processing. This strategy integrated effective signals from the time-frequency domain, autocorrelation domain and wavelet domain to enhance feature diversity. The core goal of this research was to break through the inherent limitations of traditional single-domain feature analysis methods. It specifically aimed to achieve accurate and efficient discrimination between underground winter bamboo shoots and bamboo rhizomes in complex soil environments. The proposed method designed a dedicated ICSE feature enhancement module with strong robustness. This module effectively compensated for the loss of key detailed information during the feature extraction process. It also significantly improved the overall capability of feature representation and signal differentiation. Residual Network (RES) was introduced to effectively alleviate the common gradient vanishing problem in the process of network training. An improved Convolutional Block Attention Module (CBAM) was embedded into the network structure to optimize feature extraction. This improved CBAM module enhanced the extraction of key signal features through dual complementary effects. These two effects included channel attention mechanism and spatial attention mechanism, which jointly improved feature selection accuracy. Meanwhile, three mature signal processing methods were adopted to conduct comprehensive multi-dimensional analysis of original time-domain signals. They were Short-Time Fourier Transform (STFT), Autocorrelation Calculation (AC) and Wavelet Transform (WT), each with unique advantages. STFT was utilized to convert original time-domain signals into time-frequency domain signals. It clearly revealed how signal frequencies change dynamically over time, providing important time-frequency information. AC was employed to deeply explore the autocorrelation characteristics of signals. It helped effectively capture the inherent internal laws and potential change rules of the signals. WT was applied to perform multi-scale decomposition of signals with high precision. It successfully extracted detailed feature information at different scales, which was crucial for distinguishing subtle signal differences. On this solid foundation, the Stacking multi-domain fusion strategy was combined to integrate valuable complementary information from the three feature domains. This integration fully utilized the advantages of each domain and avoided information waste. This effective integration significantly enhanced the ability to recognize subtle and weak signals in complex environments. Experimental results further demonstrated the good effectiveness and stability of the proposed strategy. The overall recognition accuracy of the proposed multi-domain feature fusion strategy reached 84.44% in practical tests. Specifically, the recognition accuracies for soil, winter bamboo shoot and bamboo rhizome signals were 86.67%, 83.75% and 82.92% respectively, showing stable performance. The multi-domain feature fusion strategy proposed in this paper achieves higher recognition accuracy than traditional single-domain feature analysis methods, which verifies the favorable effectiveness and reliability of the proposed method in the identification of winter bamboo shoots and bamboo rhizomes.