Alleviating labeled data scarcity: a lightweight semi-supervised network for Moso bamboo age determination

Accurate determination of Moso bamboo ( Phyllostachys edulis ) age is a critical task for efficient and sustainable bamboo forest management. However, existing methods face significant challenges: traditional manual assessment is subjective and labor-intensive, while advanced technologies like LiDAR are prohibitively expensive for widespread application. Furthermore, high-performance deep learning models, which offer a promising alternative, typically rely on large-scale labeled datasets, a resource that is particularly scarce and costly to acquire in the field of Moso bamboo. To address these limitations, we propose a lightweight, semi-supervised framework, the Dual-Color-Texture Moso Bamboo Age Decoupling Network (DCT-MBADNet). Our framework first leverages the Segment Anything Model (SAM) to isolate the bamboo culm, effectively eliminating complex background interference. A novel dual-stream feature decoupling module is then introduced to independently extract color degradation and texture evolution features, which are biologically significant indicators of bamboo age. A dynamic gating mechanism is employed to adaptively fuse these features. Simultaneously, we integrate an age-dependent dynamic threshold strategy within a Mean Teacher semi-supervised framework to synergistically utilize a small set of labeled data and a large volume of unlabeled data, thereby enhancing pseudo-label quality and model generalization. Experimental results demonstrate that our semi-supervised DCT-MBADNet achieves a test set accuracy of 89.6%, representing a 4.5% improvement over its fully supervised baseline. With a minimal parameter count of just 1.6 M, the proposed model provides a low-cost, robust, and deployable solution for precise Moso bamboo management and offers a novel paradigm for plant phenotyping analysis under data-scarce conditions.