An accurate estimation of seasonal spatiotemporal dynamics of forage phosphorus (P) content in alpine grassland is crucial for effective grassland and livestock management. In this study, we integrated Sentinel-2 multispectral imagery (MSI) with computational hyperspectral features (CHSFs) and developed a novel symbolic regression algorithm based on deep reinforcement learning and genetic programming (DRL-GP) to estimate forage P content in alpine grasslands. Using 243 field observations collected during the regreening, grass-bearing, and yellowing periods in 2023 from the Shaliu River Basin, we generated 10 CHSF images (CHSFIs) with varying spectral dispersions (1–10 nm). Our results demonstrated the following: (1) The DRL-GP-based symbolic regression model identified the optimal CHSF and spectral dispersion for each growing season, significantly enhancing estimation accuracy. (2) Forage P content estimations using the combined CHSF and DRL-GP-based symbolic regression algorithm significantly outperformed traditional methods. Compared to original spectral features, the R2 improved by 99.5%, 57.4%, and 86.2% during the regreening, grass-bearing, and yellowing periods, with corresponding MSE reductions of 84.8%, 41.5%, and 75.8% and MAE decreases of 70.7%, 57.5%, and 50.4%. Across these growing seasons, the R2 increased by 322.2%, 68.2%, and 639.8% compared to MLR, 128.9%, 97.4%, and 469.2% compared to RF, and 485.1%, 65.3%, and 231.3% compared to DNN. The MSE decreased by 31%, 82.9%, and 52.4% compared to MLR, 39.9%, 42.4%, and 31.4% compared to RF, and 84.5%, 73.4%, and 81.9% compared to DNN. The MAE decreased by 32.6%, 67%, and 44.2% compared to MLR, 42.6%, 47.6%, and 37.9% compared to RF, and 60.2%, 50%, and 56.3% compared to DNN. (3) Proximity to the water system notably influenced forage P variation, with the highest increases observed within 1–2 km of water sources. These findings provide critical insights for optimizing grassland management and improving livestock productivity.
