Identification of the optimal phenological periods for summer maize yield prediction using UAV-based multispectral data

1. A structured framework to identify the optimal phenological periods for summer maize yield prediction using UAV-based multispectral data was proposed. 2. Tasseling stage is the earliest suitable period for maize yield prediction. 3. Integration of data from the tasseling, silking, and dough stages was recommended in contrast to mono-temporal or full-temporal data. Timely and accurate forecasting of crop yields is critical for food management and trade. However, limited research has explored the impact of integrating crop phenotypic parameters (CPPs) with unmanned aerial vehicle (UAV) data across different phenological stages on maize yield prediction. Extent to which multi-temporal data, as opposed to mono-temporal data, enhances the accuracy and reliability of yield projections has yet to be systematically investigated. To attain the balance between accuracy and cost in crop yield estimation, this study proposed a structured framework to identify the optimal phenological periods for summer maize yield prediction using UAV-based multispectral data. Three classical methods including the custom mean decrease accuracy (C-MDA), optimal parameters-based geographical detector (OPGD), and grey relational analysis (GRA), were firstly used to sort and screen both CPPs and vegetation indices (VIs) derived from UAV-based information over six growth stages. Ridge regression models based on multi-temporal data combinations and mono-temporal data were established, respectively, whose performance in yield prediction were compared to identify the optimal phenological stages and the corresponding key factors. Our results showed that the C-MDA exhibited a pronounced superiority in factor screening ranking compared to OPGD and GRA. Green normalized difference vegetation index (GNDVI), normalized difference vegetation index (NDVI), and normalized difference red edge index (NDRE) emerged as the top-performing VIs, while the leaf area index (LAI) and above ground biomass (AGB) proved to be the most effective CPPs. When predicting yield using only mono-temporal data, the dough stage delivered the highest predictive accuracy ( R 2=0.871, RMSE=0.407 t ha -1 ), while the tasseling stage achieved yield estimates with acceptable precision ( R 2=0.810, RMSE=0.493 t ha -1 ) the earliest. In contrast, the integration of UAV data from different crop growth stages markedly enhanced the accuracy of yield estimation. Combinations of data from the tasseling, silking, and dough stages was recommended ( R 2=0.942, RMSE=0.291 t ha -1 ). These findings indicate that precise estimation of maize yields in smallholder fields may be attainable, presenting both substantial theoretical insights and practical benefits for the advancement of precision agriculture.