Using field hyperspectral data to predict cotton yield reduction after hail damage

Abstract

Rapid and accurate prediction of yield reduction after a hailstorm is essential for insurance companies for developing compensation standards and farmers to take appropriate post-damage management measures. Here, six (I-VI) hail damage simulation treatments with four damage levels (0%, 30%, 60%, and 90%) each time (for modeling) and natural hailstorm tracking experiment (for model verification) were conducted between 2018 and 2019 to explore the feasibility of yield prediction at field scale in different damage stages according to the reflectance of spectral characteristic bands (RSCB). The performance of models was analyzed with three regression methods (partial least squares regression (PLSR), support vector regression, and backpropagation neural network) based on RSCB and vegetation indices. The model developed at different damage stages showed higher accuracy in predicting yield reduction than a general model developed over the whole vegetation period. The yield prediction model based on RSCB showed better performance compared to that based on the vegetation indices, and overall, PLSR performed the best. A general model was developed for pre-bloom hail damage, whereas three models were developed for damage stage IV (accumulated growing degree days (AGDD) ranging between 975.9 °C·d and 1291.6 °C·d), V (AGDD ranging between 1291.6 °C·d and 1550.0 °C·d) and VI (AGDD greater than 1550.0 °C·d), respectively for post-bloom hail damage. Employing RSCB at different damage stages is a viable approach to predicting cotton yield reduction ascribed to hail damage.