Climate change is intensifying the frequency and severity of extreme weather events, posing significant challenges to crop productivity and agroclimatic management in subtropical regions. However, quantitative insights into how different cropping systems respond to climate extremes remain limited. In this study, crop net primary productivity (CNPP) of two representative cropping systems, early–late rice (ER–LR) and dry rapeseed–sweet potato (DR–SP), was analyzed in Pingxiang, a typical subtropical agricultural region of China. Nineteen extreme temperature and precipitation indices were evaluated using an integrated Trend–Prediction–Sensitivity–Threshold (TPST) framework combining statistical and machine learning approaches. CNPP exhibited an upward trend (slope = 4.29 g C m⁻² yr⁻¹) from 2000 to 2023, with ER–LR showing faster growth (slope = 4.54 g C m⁻² yr⁻¹) and higher stability (high-volatility area: 1.25%) than DR–SP (slope = 4.11 g C m⁻² yr⁻¹; 4.94%). Temperature extremes were the dominant drivers, exhibiting nonlinear responses with threshold effects. DR–SP was more climate-sensitive, while ER–LR showed greater tolerance, highlighting the role of cropping systems in enhancing resilience. The TPST framework provides a transferable approach for assessing agroecosystem productivity responses to climate extremes and supports climate-resilient cropland management in subtropical regions.