科研进展

【论文题目】Estimating potato nitrogen content using active transfer learning based on UAV hyperspectral imagery

【作者】Hang Yin, Haibo Yang, Fei Li, Yuncai Hu, Kang Yu

【Abstract】Significant progress has been made in estimating plant nitrogen content (PNC) using remote sensing data, driven by advancements in machine learning models. However, external factors such as interannual variability, soil heterogeneity, and fertilization management often influence model stability across different growing seasons. Transfer learning has gained attention as an effective strategy to enhance model generalization, yet the domain adaptability of transferred samples remains a pressing challenge that warrants further investigation. Therefore, this study integrated active learning with transfer learning to identify the most informative transferable samples, thereby enhancing model generalization across heterogeneous growth environments. Spectral data of potato plants were collected using uncrewed aerial vehicles (UAVs)-based hyperspectral imaging at three locations from 2018 to 2021. The transferability of an unsupervised transfer method [transfer component analysis (TCAs)] and two sample-based transfer methods (random sample update and active learning sample update) combined with partial least squares regression (PLSR) was compared across different years. The results showed that both sample-based transfer methods outperformed the unsupervised transfer method across different years. The random sample update required 35% of the target dataset to reach moderate accuracy (R²=0.40-0.70), whereas the active learning sample update attained superior performance (R²=0.48-0.75) using only 5-6 high-residual samples. This approach effectively minimized latent space misalignment within the PLSR model, thereby enhancing the model’s transferability to unseen datasets. Furthermore, the active learning sample update method identified key nitrogen-sensitive spectral bands, primarily within the 520-570 nm and 680-730 nm ranges. Thus, our study suggests that the active learning sample update could significantly advance PNC estimation and support scalable PNC monitoring using future hyperspectral satellite missions.

【Keywords】Active learning; Hyperspectral; Nitrogen; Potato; Transfer learning