Rolling prediction of summer maize yield fusing multi-source data and extremely randomized trees model

Conventional single-stage static models cannot fully capture the cumulative effects of meteorological factors and canopy growth over the growing season. It is often required to accurately and rapidly predict summer maize yield. In this study, a rolling prediction framework was developed to integrate field meteorological observations, canopy hyperspectral information, and the extremely randomized trees (ERT) model. Yield estimates were dynamically accumulated as growth-stage information. A relatively reliable prediction window was identified to improve prediction accuracy and stability for yield assessment of summer maize. Field experiments were conducted with staggered sowing dates at the Gucheng National Agro-meteorological Field Scientific Observation and Research Station in Hebei Province, China, from 2018 to 2023. Meteorological and canopy hyperspectral data was collected at six growth status, including the three-leaf, seven-leaf, jointing, tasseling, milk, and maturity stages. Furthermore, 13 vegetation indices were calculated from canopy hyperspectral reflectance to characterize canopy structure, chlorophyll status, water condition, and photosynthetic activity. Two strategies were constructed for prediction: Single-Stage Prediction (SSP) and Rolling Prediction (RP). A nested cross-validation framework was adopted for the two-stage feature selection using variance inflation factor (VIF) and feature importance screening via node impurity reduction. The 2024 field experiment data was further used for validation. The results showed that: 1) The predictors differed between the SSP and RP models, but the number of retained features was similar. The SSP models retained 6-8 features, whereas the RP models retained 7-9 features, indicating that the two-stage feature selection strategy effectively reduced feature redundancy for the predictive information. 2) The RP models shared better prediction performance than the SSP ones. On the test set, the coefficient of determination (R²) of the RP models reached 0.88 with a root mean square error (RMSE) of 636.05 kg/hm² and a mean absolute percentage error (MAPE) of 4.70%. From the tasseling to the maturity stage, rolling integration of earlier meteorological and later canopy spectra was reduced the prediction fluctuation caused by single-stage modelling. Moreover, 2024 data validation showed that the RP model was achieved in more stable performance after rolling to the milk and maturity stages, especially for the maturity stage with an R² of 0.88, an RMSE of 616.34 kg/hm², and a MAPE of 6.18%. The milk-to-maturity period was a relatively reliable window to predict the field-scale summer maize yield. 3) Feature importance analysis showed stage-dependent patterns. Before the maturity stage, yield prediction was dominated by temperature, precipitation, and their cumulative effects, which were closely related to source–sink establishment and dry matter accumulation. At the maturity stage, the importance of remote sensing vegetation indices, including the MERIS terrestrial chlorophyll index (MTCI) and red edge position index (REP), increased again, because leaf senescence caused chlorophyll degradation and canopy water loss, thereby altering canopy spectral responses. The rolling prediction framework with multi-source data and ERT model improved the accuracy and stability of summer maize yield prediction. The available stage-wise yield estimates can be updated after evaluation. The milk-to-maturity period was identified as a relatively stable window for final yield assessment, whereas earlier predictions were more suitable for preliminary risk warning with meteorological and canopy spectra.