Transfer learning for enhancing the generality of leaf spectroscopic models in estimating crop foliar nutrients across growth stages

• Enhanced foliar nutrient prediction using transfer learning across growth stages. • TrAdaBoostR2 outperforms in comparing instance-, feature-, and parameter-based methods. • Leveraging hyperspectral data absorption features improves nutrient model generality. China, despite being a leading producer of potatoes, has a potato yield below the global average, primarily due to inefficient nutrient management practices. Remote sensing provides a non-invasive and large-scale approach to monitor crop nutrient status, offering an efficient alternative to traditional plant tissue analysis. However, the generalization of foliar nutrient models is often constrained by factors such as growth stages and planting cultivars. Transfer learning offers a powerful solution by utilizing knowledge acquired from one task to enhance performance in related one, addressing challenges in model generalizability. Here, we investigated the potential of integrating various transfer learning techniques with partial least squares regression (PLSR) for retrieving three key potato foliar nutrients (nitrogen, phosphorus and potassium) across five growth stages (emergence, tuber initiation, early tuber bulking, mid-tuber bulking and tuber maturation). Three categories of transfer learning techniques were examined: 1) instance-based, including PLSR-KMM (kernel mean matching) and PLSR-TrAdaBoostR2 (transfer adaptive boosting for regression); 2) feature-based, including PLSR-TCA (transfer component analysis); and 3) parameter-based, including PLSR-parameter-based. We found that: 1) The combination of transfer learning techniques with PLSR could generally enhance the model transferability across growth stages, with a decrease in the normalized root mean squared error (nRMSE of 1–10 % for nitrogen, 3–60 % for phosphorous, and 1–15 % potassium; 2) The ranking of transfer learning techniques for improving model generalizability was: PLSR-TrAdaBootR2 > PLSR-parameter based > PLSR-recalibrated > PLSR-TCA > PLSR-KMM; 3) Foliar nitrogen demonstrated the highest transferability, followed by potassium and phosphorus; 4) PLSR models integrated with transfer learning techniques more effectively leveraged the absorption features of foliar biochemistry (e.g., chlorophyll, water and dry matters) to predict nutrients.