<p>Detecting Ochotona curzoniae in natural scenes is challenging due to their small pixel ratio and indistinct visual features. To improve object-detection accuracy, we designed an information-interactive bidirectional-feature-fusion model that embeds the Bidirectional Adaptive Feature Aggregation Network (BAFAN) into Yolov8. This design facilitates two-way information exchange between high-level semantic features and low-level spatial details. First, BAFAN adopts both bottom-up and top-down feature fusion paths with residual connections to mitigate the loss of shallow detail and ensure semantic-detailed feature complementation. To further enhance feature discrimination, an adaptive weighting mechanism is employed to assign appropriate importance to multiscale features. Second, to enhance the model’s feature-extraction capability for small objects, we incorporate Spatial-to-Depth Convolution (SPD-Conv) into the backbone network. SPD-Conv divides feature maps into multiscale subgraphs via spaced spatial sampling and encodes spatial detail into the channel dimension using non-stride convolution, thereby enhancing the representation of spatial details. Third, we propose CSFAM (CBAM-SPPF Fusion Attention Modules), a dual-attention mechanism that integrates spatial and channel attention with global pooling. It generates attention weights through a shared fully connected layer to highlight critical regions and suppress redundant information. Maxpooling is also used to capture richer contextual cues. Experiments on the Ochotona curzoniae dataset demonstrate that BC-YOLO surpasses the baseline model by 1.2% in mAP@0.5, 2.2% in mAP@0.5:0.95, 2.4% in Precision, and 1.9% in Recall. Additionally, to assess the model’s generalization ability, experiments were conducted on the NWPU VHR-10 dataset.</p>