Predicting grain yield of maize using a new multispectral-based canopy volumetric vegetation index

Abstract

Accurately predicting agricultural yields is crucial for developing adaptative strategies to ensure food security. Unmanned aerial vehicle (UAV) remote sensing equipped with portable multispectral sensors are commonly applied to acquire high temporal and spatial resolutions of remote sensing data. The vegetation indices (VIs) extracted from multispectral images are conducted for agricultural yield prediction. However, existing VIs often suffered from saturation problems when the canopy coverage is high. Integrating UAV-derived canopy height data with spectral indices holds the potential to solve saturation problem. However, this method is still at the infant stage and requires further validation. Here, we have newly proposed a multispectral-based canopy volumetric vegetation index (MSCVI) that integrates RGB-based volumetric index (VCI) and multispectral images derived VIs from UAV platform for predicting irrigated maize yields for three years (2019, 2020, and 2021). To test the stability of the proposed method, the maize was well managed and different levels of fertilizers were applied in each plot. The results using regression analysis showed the MSCVI outperformed the single adoption of VIs and VCI, and the MSCVI at reproductive growth stages was more strongly correlated with maize yields. Two commonly applied machine learning approaches: backpropagation neural network (BP) and random forest (RF) were applied for predicting maize yield. The R2 between actual maize yield and predicted maize yield using BP increased from 0.81 to 0.86 (RMSE decreased from 0.93 to 0.67 t/ha). The R2 between actual maize yield and predicted maize yield using RF increased from 0.91 to 0.94 (RMSE decreased from 0.65 to 0.42 t/ha). The robustness of the proposed model was further evaluated using data updating strategies, and results implied that the models was stable across sensors and different years. Overall, this study revealed the proposed MSCVI obtain high potential for predicting agricultural yields, and the proposed model was robust and stable when tested using data updating strategy.