Tang Minmeng, Sadowski Dennis Lee, Peng Chen, Vougioukas Stavros G, Klever Brandon, Khalsa Sat Darshan S, Brown Patrick H, Jin Yufang
Department of Land, Air, and Water Resources, University of California, Davis, Davis, CA, United States.
Department of Biological and Agricultural Engineering, University of California, Davis, Davis, CA, United States.
Front Plant Sci. 2023 Feb 15;14:1070699. doi: 10.3389/fpls.2023.1070699. eCollection 2023.
Estimating and understanding the yield variability within an individual field is critical for precision agriculture resource management of high value tree crops. Recent advancements in sensor technologies and machine learning make it possible to monitor orchards at very high spatial resolution and estimate yield at individual tree level.
This study evaluates the potential of utilizing deep learning methods to predict tree-level almond yield with multi-spectral imagery. We focused on an almond orchard with the 'Independence' cultivar in California, where individual tree harvesting and yield monitoring was conducted for ~2,000 trees and summer aerial imagery at 30cm was acquired for four spectral bands in 2021. We developed a Convolutional Neural Network (CNN) model with a spatial attention module to take the multi-spectral reflectance imagery directly for almond fresh weight estimation at the tree level.
The deep learning model was shown to predict the tree level yield very well, with a R2 of 0.96 (±0.002) and Normalized Root Mean Square Error (NRMSE) of 6.6% (±0.2%), based on 5-fold cross validation. The CNN estimation captured well the patterns of yield variation between orchard rows, along the transects, and from tree to tree, when compared to the harvest data. The reflectance at the red edge band was found to play the most important role in the CNN yield estimation.
This study demonstrates the significant improvement of deep learning over traditional linear regression and machine learning methods for accurate and robust tree level yield estimation, highlighting the potential for data-driven site-specific resource management to ensure agriculture sustainability.
估计和了解单个田块内的产量变异性对于高价值树木作物的精准农业资源管理至关重要。传感器技术和机器学习的最新进展使得以非常高的空间分辨率监测果园并在单株水平上估计产量成为可能。
本研究评估了利用深度学习方法通过多光谱图像预测单株杏仁产量的潜力。我们聚焦于加利福尼亚州一个种植‘独立’品种的杏仁果园，在那里对约2000棵树进行了单株收获和产量监测，并于2021年获取了30厘米分辨率的夏季航空图像，涵盖四个光谱波段。我们开发了一个带有空间注意力模块的卷积神经网络（CNN）模型，直接利用多光谱反射图像来估计单株杏仁的鲜重。
基于五折交叉验证，深度学习模型被证明能很好地预测单株产量，决定系数R2为0.96（±0.002），归一化均方根误差（NRMSE）为6.6%（±0.2%）。与收获数据相比，CNN估计很好地捕捉了果园行间、沿样带以及单株之间的产量变化模式。发现红边波段的反射率在CNN产量估计中起最重要作用。
本研究表明，与传统线性回归和机器学习方法相比，深度学习在准确且稳健的单株产量估计方面有显著改进，凸显了数据驱动的特定地点资源管理对确保农业可持续性的潜力。
