【摘要】 【目的】水稻籽粒蛋白质含量（grain protein content,GPC）是衡量稻米品质和商品价值的重要指标。建立快速、无损的水稻GPC估测方法，旨在为作物智慧育种提供理论依据和技术支持。【方法】采用无人机搭载RGB相机和多光谱相机，于2022—2023年获取522份水稻育种材料抽穗-成熟期的RGB和多光谱影像及实测GPC数据。利用Gram-Schmidt图像融合方法对RGB和多光谱影像进行处理得到融合影像，并结合基于原始多光谱图像提取的光谱特征和纹理特征，采用随机森林（random forest,RF）、极限梯度提升机（extreme gradient boosting,XGBoost）、梯度提升回归（gradient boosting regression,GBR）3种机器学习回归算法构建GPC估测模型。【结果】RGB影像的红波段包含更丰富的图像信息，经过该波段融合后的植被指数与GPC的相关性均高于由原始多光谱影像计算的植被指数。均值纹理（Mean）在纹理指数构建中出现频率最高（占比63.16%），其中MEA₅₆₀-MEA₈₄₀指数与不同类型水稻的GPC具有一定的相关性（淮安常规粳稻：|r²|=0.28；如皋杂交粳稻：|r²|=0.20）。以多光谱图像特征、纹理特征和融合图像特征作为输入参数组合构建的水稻GPC估测模型，在抽穗期（R2 cal=0.64）和成熟期（R2 cal=0.70）的精度高于灌浆期模型（R2 cal=0.53）。相较于使用原始影像特征，结合融合影像特征提高了GPC的估测精度（ΔR2 cal=0.08-0.26）。RF构建的年际模型精度高于XGBoost和GBR模型（RF:R2 val=0.74,RMSE=0.21%;XGBoost:R2 val=0.58,RMSE=0.23%;GBR:R2 val=0.42,RMSE=0.23%）。【结论】结合无人机影像融合技术和机器学习方法能有效提高水稻育种材料GPC的估测精度，研究结果可为大规模水稻品质参数精准估算提供理论参考和有效途径。更多还原

【Abstract】 【Objective】Grain protein content（GPC） is a crucial indicator for evaluating rice quality and its commercial value. Establishing a rapid and non-destructive method for estimating rice GPC was established, so as to provide theoretical foundations and technical support for smart breeding and precision crop management. 【Method】 This study employed a drone equipped with both an RGB camera and a multispectral camera to collect RGB and multispectral imagery, along with ground-measured grain protein content（GPC） data, from the heading to maturity stages of 522 rice breeding material accessions from 2022 to 2023. The Gram-Schmidt image fusion method was applied to process the RGB and multispectral images for generating fused images. Spectral and texture features extracted from the original multispectral images were combined with fused image features, and three machine learning regression algorithms—Random Forest（RF）, Extreme Gradient Boosting（XGBoost）, and Gradient Boosting Regression（GBR）—were employed to construct GPC estimation models. 【Result】The R-band of the RGB images contained richer image information. Vegetation indices derived from the fused R-band exhibited higher correlations with GPC than those calculated from the original multispectral data. The mean texture（Mean） appeared most frequently in texture index construction（accounting for 63.16%）, with the MEA₅₆₀-MEA₈₄₀ index showing certain correlations with GPC across different rice types（Huaian conventional japonica: |r²|=0.28; Rugao hybrid japonica: |r²|=0.20）. Using a combination of multispectral image features, texture features, and fused image features as input parameters, the GPC estimation models for rice breeding materials achieved higher accuracy at the heading stage（R2 cal=0.64） and maturity stage（R2 cal=0.70） than at the filling stage model（R2 cal=0.53）. Incorporating fused image features improved GPC estimation accuracy（ΔR2 cal=0.08-0.26） over using original image features. The interannual model of RF outperformed those of XGBoost and GBR in accuracy（RF: R2 val=0.74, RMSE=0.21%; XGBoost: R2 val=0.58, RMSE=0.23%; GBR: R2 val=0.42, RMSE=0.23%）. 【Conclusion】 The integration of UAV image fusion technique and machine learning methods could effectively enhance the estimation accuracy of the grain protein content（GPC） in rice breeding materials. These findings provided a theoretical reference and practical approaches for the precise estimation of rice quality parameters on a large scale.更多还原