Yoon Sunghyun, Jo Jung Su, Kim Steven B, Sim Ha Seon, Kim Sung Kyeom, Kim Dong Sub
Department of Artificial Intelligence, Kongju National University, Cheonan 31080, South Korea.
Department of Horticultural Science, College of Agricultural & Life Science, Kyungpook National University, Daegu 41566, South Korea.
Heliyon. 2023 Mar 13;9(3):e14546. doi: 10.1016/j.heliyon.2023.e14546. eCollection 2023 Mar.
The receptacle of strawberry is a more direct part than the flower for predicting yield as they eventually become fruits. Thus, we tried to predict the yield by combining an AI technique for receptacle detection in images and statistical analysis on the relationship between the number of receptacles detected and the strawberry yield over a period of time. Five major cultivars were cultivated to consider the cultivar characteristics and environmental factors for two years were collected to consider the climate difference. Faster R-CNN based object detector was used to estimate the number of receptacles per strawberry plant in given two-dimensional images, which achieved a mAP of 0.6587 for our dataset. However, not all receptacles appear on the two-dimensional images, and Bayesian analysis was used to model the uncertainty associated with the number of receptacles missed by the AI. After estimating the probability of fruiting per receptacle, prediction models for the total strawberry yield at the end of harvest season were evaluated. Even though the detection accuracy was not perfect, the results indicated that counting the receptacles by object detection and estimating the probability of fruiting per receptacle by Bayesian modeling are more useful for predicting the total yield per plant than knowing its cumulative yield during the first month.
草莓的花托比花朵更能直接预测产量，因为花托最终会发育成果实。因此，我们尝试通过结合一种用于图像中花托检测的人工智能技术以及对检测到的花托数量与一段时间内草莓产量之间关系的统计分析来预测产量。种植了五个主要品种以考虑品种特性，并收集了两年的环境因素数据以考虑气候差异。使用基于Faster R-CNN的目标检测器来估计给定二维图像中每株草莓的花托数量，对于我们的数据集，该检测器的平均精度均值（mAP）达到了0.6587。然而，并非所有花托都会出现在二维图像上，因此使用贝叶斯分析对与人工智能遗漏的花托数量相关的不确定性进行建模。在估计每个花托结果的概率后，对收获季节结束时草莓总产量的预测模型进行了评估。尽管检测精度并不完美，但结果表明通过目标检测对花托进行计数并通过贝叶斯建模估计每个花托结果的概率，比了解其第一个月的累计产量更有助于预测单株总产量。
