Abstract Crop red–green–blue (RGB) images are powerful tools in nitrogen (N) nutrition estimation. Various regression models using crop N nutrition parameters and image indices have been suggested, but their accuracy and generalization performance for N estimation have not been thoroughly evaluated. In this study, a commercial digital camera was used to capture rice canopy RGB images in a 2-year field experiment, and three regression methods (simple nonlinear regression, SNR; backpropagation neural network, BPNN; and random forest regression, RF) were used for rice shoot dry matter (DM), N accumulation (NA), and leaf area index (LAI) estimation. A repeated random subsampling validation method was performed 1000 times on all three regression methods for the evaluation of model performance and stability. The RF regression models had the highest accuracy for the validation dataset, with average testing prediction accuracy (ATPA) of 80.17%, 79.44%, and 81.82% for DM, LAI, and NA estimation, respectively, followed by BPNN and SNR models. According to the distribution of ATPA in 1000-time calculations, the highest standard deviation (SD) and interval range (5%–95%) of ATPA was observed in BPNN models, which indicated that the BPNN model was most susceptible to dataset splitting. The lower SD and interval range of ATPA were followed by RF and SNR models, which indicated that the RF and SNR models were less affected by dataset splitting and were able to produce robust regression models consistently. In conclusion, the ensemble algorithm of the RF model effectively prevents overfitting when dealing with different dataset segmentations; thus, the RF model has strong generalization performance. A combination of digital imagery and appropriate machine learning methods facilitates convenient and reliable estimation of crop N nutrition.

中文翻译：

---

用RGB图像和机器学习方法估计水稻氮营养

摘要 作物红-绿-蓝 (RGB) 图像是氮 (N) 营养评估的有力工具。已经提出了使用作物 N 营养参数和图像指数的各种回归模型，但它们对 N 估计的准确性和泛化性能尚未得到彻底评估。在这项研究中，使用商用数码相机在为期 2 年的田间试验中捕获水稻冠层 RGB 图像，并使用三种回归方法（简单非线性回归，SNR；反向传播神经网络，BPNN；和随机森林回归，RF）用于水稻芽干物质 (DM)、氮积累 (NA) 和叶面积指数 (LAI) 估计。对所有三种回归方法进行了 1000 次重复随机子采样验证方法，以评估模型性能和稳定性。RF 回归模型对验证数据集的准确度最高，DM、LAI 和 NA 估计的平均测试预测准确度 (ATPA) 分别为 80.17%、79.44% 和 81.82%，其次是 BPNN 和 SNR 模型。根据ATPA在1000次计算中的分布，在BPNN模型中观察到ATPA的最高标准差（SD）和区间范围（5%~95%），表明BPNN模型最容易发生数据集分裂。ATPA 较低的 SD 和区间范围之后是 RF 和 SNR 模型，这表明 RF 和 SNR 模型受数据集拆分的影响较小，并且能够始终如一地生成稳健的回归模型。综上所述，RF模型的集成算法在处理不同的数据集分割时有效地防止了过拟合；因此，RF模型具有很强的泛化性能。数字图像和适当的机器学习方法相结合，可以方便可靠地估计作物 N 营养。