Early prediction of the seed yield in winter oilseed rape based on the near-infrared reflectance of vegetation (NIRv)

Abstract

Demands for rape seeds oil rapidly increase in recent years. The empirical model based on the remotely sensed data provides an efficient approach to predict the rapeseed yield at small scale. The vegetation indices (VIs) derived from the remotely sensed data at the stage of the peak leaf area index (LAI) are usually the main determinant of the empirical model. The LAI of oilseed rape reaches the peak during the flowering stage. However, yellow flowers elevate the reflectance in the red band, making it difficult to capture the maximum LAI with the commonly-used VIs, such as the normalized difference vegetation index (NDVI). This study tested the hypothesis that the near-infrared reflectance of vegetation (NIRv) may be more sensitive to the LAI in the oilseed rape throughout different growth stages, particularly in the flowering stage, and thus may provide more accurate early prediction of the rapeseed yield as well as the above-ground dry mass (DM). In addition, a random forest (RF) regression model based on the multi-stage NIRv was built to predict the rapeseed yield. Three small-plot experiments with different nitrogen and potassium fertilizer treatments were conducted at two sites in Hubei province, China. NIRv and the studied VIs were derived from multi-spectral images captured by an unmanned aerial vehicle. NIRv was calculated with the DN values (NIRv_DN) and the reflectance values (NIRv_refl) of the near-infrared band separately. Results demonstrated that NIRv_DN and NIRv_refl were strongly and consistently correlated with the LAI before the flowering stage and in the flowering stage. In comparison, the relationships between the studied VIs and LAI in the flowering stage deviated from the relationships of all growth stages. The NIRv features and VIs were all significantly correlated with the above-ground DM, but the DM estimation by NIRv_DN had a lower RMSE value. The NIRv_DN and NIRv_refl in the flowering stage showed the strongest correlation with the rapeseed yield (Cali_ R² = 0.74 and Vali_RMSE = 389.84 kg/ha for NIRv_DN; Cali_ R² = 0.73 and Vali_RMSE = 410.85 kg/ha for NIRv_refl). The NIRv_DN, NIRv_refl, and the soil adjusted vegetation index (SAVI) in the budding stage were also strongly correlated with the rapeseed yield. The variable importance score derived from the RF model corroborated the significant contribution of the NIRv in the flowering and budding stage to the yield prediction. The RF model using the NIRv_DN in the flowering and budding stage achieved the high accuracy of the rapeseed yield prediction (Cali_ R² = 0.91, Cali_RMSE = 248.48 kg/ha, and Vali_RMSE = 282.44 kg/ha). Results from this research demonstrated the great potential of the NIRv in the flowering stage to provide the accurate early prediction of the rapeseed yield.

Plant phenotyping relevance

UAVマルチスペクトル画像からNIRvを抽出し、LAI・乾物量・収量を推定する手法を比較・検証しており、植物形質取得と予測ワークフローが研究の中心である。

Code and data availability

公開本文の所在を確認できませんでした。非公開または購読が必要な可能性があります。

No evidence-backed public reproduction asset is currently recorded.