Synergistic use of stay-green traits and UAV multispectral information in improving maize yield estimation with the random forest regression algorithm

Abstract

The timely and accurate estimation of maize yield in the field is critical for developing agricultural management strategies and ensuring food security. Most crop yield estimation models are based on absolute indices, such as spectral and textural indices, yet few studies have considered the stability of relative indices in the time dimension. To address this issue, a novel relative index, the Stay-Green Index (SGI), was extracted to characterize maize growth from the NDVI time series of Unmanned Aerial Vehicle (UAV) multispectral images. We used the Random Forest Regression (RFR) framework and a set of vegetation indices (VIs) to build the maize yield estimation model. The effects of SGI as an additional variable on yield estimation were examined at different growth stages. We finally evaluated the performance of our yield estimation model incorporating SGI. Our results showed that: (1) the correlation coefficients between VIs and the actual yield were relatively low at the silking (R1) and blister (R2) stages, close to 0, while the results were relatively high at the dough (R4), denting (R5), and maturity (R6) stages, close to 0.69; (2) the proposed SGI was significantly correlated with the actual yield; and (3) our yield estimation model with SGI can greatly improve the accuracy of yield estimation models at the R1 and R2 stages, while the performance was maintained at the R4, R5, and R6 stages. Overall, this study highlights that the model using the RFR algorithm combining SGI and VIs can improve the within-season yield estimation for maize.