Farm-Scale Estimation of Key Maize Phenological Stages Using HLS Time Series and Thermal-Window Constraints in Heilongjiang Province, China

Accurate estimation of key maize phenological stages at the farm scale can support crop management, risk assessment, and yield prediction. Under small-sample and cross-farm conditions, however, remote-sensing time series are often affected by observation gaps, mixed pixels, and stage-dependent canopy responses, while fixed accumulated-heat thresholds may vary across farms and years. This study tested an interpretable HLS-AGDD framework for estimating the jointing, tasseling, silking, and physiological maturity of maize. The analysis was based on 12 maize sample-year units from Jianshe Farm and Junchuan Farm in Heilongjiang Province, China, during 2021–2022. NDVI, EVI2, and LSWI were extracted using 30 m and 60 m buffer radii, and AGDD was used to construct thermal-time axes from the sowing date and from 1 May. In the internal assessment, the main HLS-AGDD scheme showed low apparent errors, whereas cross-validation indicated greater uncertainty, especially when an entire year was withheld. The farm-group calendar baseline produced lower errors than the HLS-AGDD framework under the validation settings, showing that farm-specific historical dates were a strong reference in the present two-farm, two-year dataset. AGDD was mainly useful for excluding phenologically unlikely dates and limiting the search for each stage to a narrower candidate period. Within this period, HLS vegetation-index time-series features provided canopy-development information for candidate-date selection. Physiological maturity required a separate treatment because late-season canopy senescence produced delayed estimates when the same scoring rule was used for all stages. In the internal version comparison, identifying the first sustained senescence signal reduced this delay. Overall, this study provides a transparent HLS-AGDD workflow with explicitly defined thermal windows, stage-specific detection rules, and sample-level error reporting. The results also show that farm-specific calendar baselines should be included in small-sample farm-scale phenology evaluation when historical records are available.