Discharge fertilizer rate prediction across fertilizer type for a bivariate fertilization system based on transfer learning

Accurate prediction of discharge fertilizer rate (DFR) is essential for achieving precise control in bivariate fertilizer applicators (BFA). However, most existing prediction models calibrated for specific fertilizer types ignore variations in material properties, limiting their generalizability and necessitating time-consuming recalibration for different fertilizers. Transfer learning (TL) offers a promising solution by leveraging knowledge from a source domain (D_S) to improve performance in a related target domain (D_T). This study investigated the cross-fertilizer DFR prediction in a BFA system using two transfer scenarios: from a compound fertilizer (Stanley) to another (Sakefu), and from Stanley to urea. Decision tree regression (DTR) serves as the base model. Five experimental configurations were compared: direct source domain prediction (DTR- D_S); partial target domain prediction with 20% data (DTR- D_T(20\text% ) ; joint training (DTR-Recalibration); instance-based TL using transfer adaptive boosting for regression (TrAdaBoost.R2) (DTR-TrAdaBoost.R2); and full target domain prediction (DTR-D_T). The results demonstrate that: 1) DTR-Recalibration is most effective for Stanley to Sakefu, improving R² by 2.4% and reducing nRMSE by 12.5%, while DTR-TrAdaBoost.R2 performs best for Stanley to urea, increasing R² by 14.3% and reducing nRMSE by 21.5%; 2) differences in material properties significantly affect TL performance; the ranking of the importance of material characteristics is as follows: the combined difference in equivalent diameter and repose angle > bulk density > sphericity ratio; and 3) high-weight instances in TL are predominantly concentrated at medium operational intensity levels, and the stability of the discharge ratio ( Q_target/Q_source ) critically influences transfer effectiveness. These findings provide practical and theoretical insights for cross-fertilizer model transfer, enabling efficient adaptation of DFR prediction models with minimal target data and reducing calibration effort in precision agriculture applications.