A high-resolution machine learning and multi-source remote sensing approach for estimating net primary productivity in campus green spaces

This research proposes an innovative framework that integrates Machine Learning classification, the Carnegie-Ames-Stanford Approach (CASA), unmanned aerial vehicles (UAV), and multi-source data to estimate the net primary productivity (NPP) of campus green spaces. Green space was extracted using Random Forest with an overall accuracy of 94.6% (Kappa = 0.919). The fraction of Photosynthetically Active Radiation (FPAR) was derived from UAV multispectral vegetation indices, while temperature, solar radiation, and the Water Stress Factor (WSF) estimated from Sentinel 2 imagery were integrated into CASA to model NPP. Validation showed that our approach offered higher R² and lower RMSE compared to traditional CASA. Across two seasons, the model achieved R² values of 0.68-0.79, with RMSE of 11.74-17.01 gC·m^-2·month^-1 and RRMSE of 15.63-22.65%. Seasonal dynamics and human activities explained the observed variation in R², RMSE, and RRMSE. Our scalable approach suggests that campus green spaces have the potential to serve as contributors to urban carbon sinks. This method has strong potential to inform sustainable urban planning, carbon offset programs, and nature-based climate solutions.