Abstract: First-arrival picking is a crucial step in seismic data processing, as its accuracy directly affects velocity model building and static correction. Although conventional CNN-based deep learning methods have achieved remarkable results in first-arrival picking, their performance degrades in a survey with complex surface conditions, e.g. loess tableland, due to the weak energy of first arrivals and strong noises. To address this issue, we propose a deep learning-based first-arrival picking method that integrates multi-data fusion with an adaptive weighted hybrid loss function. To enhance the robustness of the method, seismic, offset, and elevation data are integrated to construct a multi-data fusion model. To enhance the accuracy of first-arrival picking, an adaptive weighting strategy is employed to optimize the combination of multiple loss functions and construct an adaptive weighted hybrid loss function, which effectively constrains the model training process. The tests on three field seismic datasets demonstrate that our method outperforms conventional methods, e.g. STA/LTA and deep semantic segmentation, in picking accuracy and noise robustness in the geologic complexity scenarios with weak first arrivals and strong noises. These results validate the effectiveness and robustness of the proposed method.