| To clarify the dynamic accumulation characteristics of aboveground biomass (AGB) of silage maize at different growth stages in the Hexi irrigation area and to improve the prediction accuracy of final AGB at maturity, silage maize experimental plots under different water and fertilizer treatments in Minle County, Zhangye City, Gansu Province, were selected as the study area. Four-band UAV multispectral images, plant height, stem diameter, leaf area, chlorophyll content, and measured AGB data were synchronously collected at the seedling, jointing, tasseling, silking, and maturity stages. The temporal variation in AGB across growth stages, AGB increments between adjacent stages, and their relationships with final AGB at maturity were analyzed. A CNN–BiLSTM–Attention model supporting variable prefix inputs was then developed to compare the early prediction performance of final AGB at maturity using different sets of observed growth stages before maturity, while the full-growth-stage input condition was used as a reference for model performance. The results showed that silage maize AGB exhibited a clear stage-specific accumulation pattern throughout the growth process. The largest AGB increment occurred from the jointing to tasseling stage, and this increment showed the strongest correlation with final AGB at maturity, indicating that this interval was the key growth period for final AGB formation. Under the early-prediction scenarios before maturity, the R² of Prefix-1 to Prefix-4 increased from 0.28 to 0.66, while the RMSE decreased from 2.14 to 1.47 t·mu?¹. Under Prefix-5, which represented the full-growth-stage input condition and served as the performance reference, R² reached 0.72 and RMSE was 1.34 t·mu?¹. Attention-weight analysis showed that the model assigned relatively high weights to the jointing and tasseling stages, which was consistent with the results of the AGB dynamic accumulation analysis. These findings indicate that multi-temporal UAV multispectral data can effectively characterize the AGB formation process of silage maize in the Hexi irrigation area. The proposed method enables prediction of final AGB at maturity using observations available at different pre-maturity growth stages, while the full-growth-stage input provides a benchmark for model performance. This approach can provide a reference for silage maize growth monitoring, yield estimation, and water and fertilizer management under irrigated agricultural conditions in arid regions. |