PKBFM: A Prior Knowledge-Based Deep Learning Framework for Bamboo Forest Mapping

Bamboo forests are widely distributed in subtropical and tropical regions and play an important role in regulating climate change and maintaining the carbon balance of ecosystems. Due to challenges such as the difficulty in obtaining field samples, the reflectance similarity between bamboo and other forests, and the difficulties in synthesizing spatial and temporal information, precise bamboo forest mapping (BFP) remains challenging. This paper proposes a novel bamboo forest mapping framework based on prior knowledge (PKBFM framework). This framework integrates the preliminary label acquisition of bamboo forests based on prior knowledge, noise label filtering, and the refined recognition of bamboo forests using an attention-based temporal-spatial network (TSNet) into a unified whole, achieving high-precision BFP while significantly reducing the reliance on field samples. The TSNet in this framework is a novel network we proposed for efficiently aggregating multi-scale spatiotemporal information, which enhances the ability to capture pixel-level spatial aggregation and phenological information in bamboo forests. Experiments conducted in two study areas, which are located respectively in Zhejiang and Hunan provinces, China, demonstrate that the PKBFM framework combined with TSNet achieves the best accuracy in bamboo mapping. Compared to other mainstream models, the F1 score improved by 0.875% to 26.518%, and the IoU increased by 1.638% to 40.520%, proving the reliability and superiority of the bamboo mapping framework we proposed. This study proposes a feasible and high-precision framework for bamboo forest mapping, which reduces the need for field samples and offers technical support for forest information monitoring.