Contents lists available at ScienceDirect Computers and Electronics in Agriculture journal homepage: www.elsevier.com/locate/compag Original papers Identification of haploid and diploid maize seeds using convolutional neural networks and a transfer learning approach Yahya Altuntaş a, ⁎ , Zafer Cömert b , Adnan Fatih Kocamaz a a İnönü University, Department of Computer Engineering Malatya, Turkey b Samsun University, Department of Software Engineering, Samsun, Turkey ARTICLE INFO Keywords: Haploid Diploid Maize Convolutional neural network Classification ABSTRACT Maize is one of the most significant grains cultivated all over the world. Doubled-haploid is an important technique in terms of advanced maize breeding, modern crop improvement and genetic programs, since this technique shortens the breeding period and increases breeding efficiency. However, the selection of the haploid seeds is a major problem of this breeding technique. This process is frequently conducted manually, and this unreliable situation leads to loss of time and labor. Inspired by the recent successes of deep transfer learning, in this study, we approached this problem as a computer vision task to provide a nondestructive, rapid and low-cost model. To achieve this objective, we adopted convolutional neural networks (CNNs) to recognize haploid and diploid maize seeds automatically through a transfer learning approach. More specifically, AlexNet, VVGNet, GoogLeNet, and ResNet were applied for this specific task. The experimental study was carried out using a new dataset consisting of 1230 haploid and 1770 diploid maize seed images. The samples in the dataset were clas- sified considering a marker-assisted selection, known as the R1-nj anthocyanin marker. To measure the success of the CNN models, we utilized several performance metrics, such as accuracy, sensitivity, specificity, quality index, and F-score derived from the confusion matrix and receiver operating characteristic curves. According to the experimental results, the CNN models ensured promising results, and we achieved the most efficient results via VGG-19. The accuracy, sensitivity, specificity, quality index, and F-score of VGG-19 were 94.22%, 94.58%, 93.97%, 94.27%, and 93.07%, respectively. Consequently, the experimental results proved that CNN models can be a useful tool in recognizing haploid maize seeds. Furthermore, we conclude that this approach is significantly superior to machine learning-based methods and conventional manual selection. 1. Introduction Maize (Zea mays L.) is one of the most significant agricultural pro- ducts used as human food, animal feed and industrial raw materials (Cerit et al., 2016). A growing world population and climate change make it necessary to develop new maize varieties that are high-yield and resistant to biotic and abiotic stress conditions like all other culti- vated plants. Achieving this goal is only possible through maize breeding programs. The first stage in maize breeding programs is to develop homozygote lines that will be parents to hybrid varieties (Cerit et al., 2016). Normally, the acquisition of homozygote lines takes a long time, approximately five to eight generations of self-cross mating by conventional methods, whereas this process can be achieved in about two to three generations in one year through the use of haploids (Prasanna et al., 2012). Haploids and doubled haploids (DH) have high importance in modern maize breeding, since this technique accelerates the breeding period and increases breeding efficiency (Chase and Nanda, 1969). These advantages of DH have led to increased interest in maize breeding and genetics in the last 20 years (Geiger, 2009). A DH is a completely homozygous line produced by doubling the haploid chromosomes (Prasanna et al., 2012). Haploids are found in nature at a very small frequency of 0.1% (Geiger et al., 2013); there- fore, they are not suitable for practical use (Charity et al., 2017). Haploids can be obtained at higher rates by using either in vitro or in vivo techniques. Most commercially available DH maize lines are ob- tained through the haploid technique in vivo while other techniques are reported to be less effective in the development of DH lines (Geiger, 2009). In vivo maternal haploid induction uses special genotypes, called inducers, as pollinators to obtain haploids at higher rates and has be- come the standard method (Charity et al., 2017). Due to currently https://doi.org/10.1016/j.compag.2019.104874 Received 9 January 2019; Received in revised form 2 May 2019; Accepted 24 June 2019 ⁎ Corresponding author at: Samsun University, Department of Software Engineering, Canik Yerleşkesi Gürgenyatak Mahallesi Merkez Sokak No:40-2/1, 55080 Canik, Samsun, Turkey. E-mail addresses: yahyaaltuntas@gmail.com (Y. Altuntaş), zcomert@samsun.edu.tr (Z. Cömert), fatih.kocamaz@inonu.edu.tr (A.F. Kocamaz). Computers and Electronics in Agriculture 163 (2019) 104874 0168-1699/ © 2019 Elsevier B.V. All rights reserved. T