Biomass estimation of World rice (Oryza sativa L.) core collection based on the convolutional neural network and digital images of canopy
Biomass estimation of World rice (Oryza sativa L.) core collection based on the convolutional neural network and digital images of canopy
复制标题
DOI：
10.1080/1343943x.2023.2210767
复制
发表时间：
2023-05-13
影响因子：
2.5
通讯作者：
Shiraiwa，Tatsuhiko
中科院分区：
文献类型：
--
作者：
Nakajima，Kota;Tanaka，Yu;Shiraiwa，Tatsuhiko
地上生物量（AGB）是衡量作物生产力的重要指标。AGB的破坏性测量成本巨大，且大多数非破坏性估算不能适用于具有不同冠层结构的品种。对AGB数据的获取不足可能限制了作物生产力的提高。最近，一种称为卷积神经网络（CNN）的深度学习技术由于其高数字图像识别能力而被应用于估计作物AGB。然而，基于cnn的AGB估计对不同品种的通用性仍不清楚。利用世界水稻核心馆藏的59个不同品种的数字图像，建立了基于cnn的水稻AGB估计方法，并对其进行了评价。在两年的时间里，我们在两个地点用商用数码相机拍摄了59个品种的12183张图像，并人工获得了相应的AGB。使用28个品种建立CNN模型，对测试数据集具有较高的准确率（R2= 0.95）。我们进一步利用31个未建立模型的品种来评估CNN模型的性能。当观测到的AGB小于924 g m−2时，CNN模型成功估计了AGB (R2= 0.87)，而当观测到的AGB大于924 g m−2时，CNN模型低估了AGB （R2= 0.02）。这种低估可以通过在进一步的研究中添加AGB更大的训练数据来改善。本研究表明，这种基于cnn的估计方法具有很强的通用性，可作为监测不同品种作物AGB的实用工具。
Above-ground biomass (AGB) is an important indicator of crop productivity. Destructive measurements of AGB incur huge costs, and most non-destructive estimations cannot be applied to diverse cultivars having different canopy architectures. This insufficient access to AGB data has potentially limited improvements in crop productivity. Recently, a deep learning technique called convolutional neural network (CNN) has been applied to estimate crop AGB due to its high capacity for digital image recognition. However, the versatility of the CNN-based AGB estimation for diverse cultivars is still unclear. We established and evaluated a CNN-based estimation method for rice AGB using digital images with 59 diverse cultivars which were mostly in World Rice Core Collection. Across two years at two locations, we took 12,183 images of 59 cultivars with commercial digital cameras and manually obtained their corresponding AGB. The CNN model was established by using 28 cultivars and showed high accuracy (R2= 0.95) to the test dataset. We further evaluated the performance of the CNN model by using 31 cultivars, which were not in the model establishment. The CNN model successfully estimated AGB when the observed AGB was lesser than 924 g m−2(R2= 0.87), whereas it underestimated AGB when the observed AGB was greater than 924 g m−2(R2= 0.02). This underestimation might be improved by adding training data with a greater AGB in further study. The present study indicates that this CNN-based estimation method is highly versatile and could be a practical tool for monitoring crop AGB in diverse cultivars.