As macaques serve as an ideal biological model for human disease simulation and pharmaceutical testing, accurate analysis of their behaviors and postures is crucial. Traditional methods for estimating macaque postures present a number of challenges, including the inefficiency of manual observation and the instability of sensor-based systems. In recent years, advances in deep learning have provided new possibilities for pose estimation in non-wearable devices. Lite-HRNet reduces the consumption of computational resources, but it loses substantial spatial information when dealing with complex environments. HRFormer increases the ability to handle multi-scale variation but its higher computational requirements limit its use in resource-constrained realistic application scenarios. Addressing these issues, we propose a novel lightweight macaque pose estimation model, named SPANet, specifically designed for natural environments. SPANet combines the multi-resolution parallel structure of HRNet with the lightweight design philosophy of ShuffleNetV2. Through path aggregation weighting and window pyramid self-attention mechanisms, it optimizes the processing of spatial information and computational efficiency. On the public MacaquePose dataset, SPANet-Tiny outperforms Lite-HRNet by improving the Average Precision (AP) by 2.7%, with 57.3% fewer parameters and 16.1% less computational cost. Compared to HRFormer-Tiny, SPANet-Tiny achieves comparable performance with 69.9% fewer parameters and 81.3% less computational cost. SPANet-Base reaches state-of-the-art performance with 79.9 AP. This demonstrates that SPANet can effectively perform macaque pose estimation in natural environments while significantly reducing resource consumption, providing robust support for macaque welfare and behavioral research.