Improving subseasonal rainfall forecasts over the middle-lower reaches of the Yangtze River: a machine learning approach using circulation predictors

Abstract Accurate subseasonal forecasts over the middle and lower reaches of the Yangtze River (MLYR) are essential for mitigating socio-economic impacts of summer rainfall anomalies. We develop a convolutional neural network (CNN) to post-process European Centre For Medium-Range Weather Forecasts subseasonal-to-seasonal forecasts by incorporating key atmospheric variables at 200 and 850 hPa (winds, geopotential height, specific humidity), which exhibit substantially higher inherent predictability than precipitation—with gh200 maintaining skill for 24–25 d, nearly three times that of precipitation. This approach improves extended boreal summer (May–September) precipitation predictions over the MLYR. The CNN-based correction consistently enhances forecast skill over lead weeks 1–4, increasing a composite index (CI; based on ACC, RMSE, MAE) by 12.3%, 4.6%, 4.2%, and 3.8%, respectively. At Week 2, the correction elevates forecasts from unusable to practically valuable, with NSE increasing from 0.297 to 0.405 and KGE from −0.666 to 0.486. Physically, CNN-derived heatmaps align with key circulation systems Jianghuai cyclone, demonstrating its capacity to capture dominant rainfall processes. Pattern correlations with observations remain high (0.89 and 0.75) in Weeks 1–2, confirming physically meaningful learning, while correct large‐scale patterns persist even at Week 4, indicating robust predictor extraction. Ablation tests further identify core predictors: total precipitation is irreplaceable during the first two weeks, while the 200 hPa zonal wind remains crucial throughout four weeks, highlighting a shift from upper‐air dynamics to lower‐level thermodynamics.