单倍体育种技术是玉米育种新方法, 该方法可有效缩短产生纯合系的周期, 提高育种效率。 该技术需首先挑选足量单倍体籽粒, 而玉米在未加人工干预时, 单倍体在混合籽粒中仅占0.05%～0.1%, 即使采用生物诱导技术, 单倍体籽粒数一般也不到籽粒总数的10%。 高速、 精准地从大量混合籽粒中挑选得到占比少于10%的单倍体籽粒, 才能够满足工程化育种需要, 而实际育种工作中挑选单倍体时常用的分子生物学、 田间形态学辨别等方法存在耗时长、 成本高、 破坏样本等缺点, 难以高效精准地得到玉米单倍体籽粒。 相关研究已经证明高油玉米的单倍体与二倍体之间具有明显含油率差异, 目前低场核磁共振技术可用于检测玉米单籽粒的含油率, 并根据含油率对单倍体进行鉴别, 但核磁共振仪存在价格贵、 维护难、 速度慢、 效率低等弱点, 现有设备完成单籽粒分选需用时4 s, 无法满足工程化育种中大量筛选的速度需求。 使用VIAVI微型近红外光谱仪能够达到0.25 s每颗的检测速度, 相比核磁共振技术速度快, 仪器价格较低, 维护方便。 使用近红外光谱仪分析技术对单倍体与二倍体籽粒进行鉴别, 可以取代核磁共振鉴别单倍体的方法。 采用近红外光谱定性鉴别单倍体籽粒虽然取得了一定效果, 但目前研究中所采集玉米品种相对较少, 研究只针对某一品种单倍体建立模型, 对该品种单倍体进行分类; 国内外尚无多品种混合单倍体鉴别相关研究, 而工程化育种亟需一种能够识别多个品种玉米单倍体的鉴别方法。 为此, 本文提出一种基于深度信念网络的多品种混合玉米籽粒单倍体鉴别方法, DBN是一种多层深度神经网络, 每层由受限玻尔兹曼机构成, 采用逐层训练策略, 可解决传统神经网络训练方法不适用于多层网络训练的问题。 对比实验结果表明使用DBN方法建立多品种单倍体鉴别模型具有较高分类性能, 能够满足玉米工程化育种精度要求。

Haploid breeding technology is a new method for maize breeding, which can effectively shorten the cycle of homozygous lines and improve the breeding efficiency. The technology needs to select enough haploid grains first, and the haploid grains only account for 0.05%～0.1% of the mixed grains without artificial intervention. Even with the biological induction technology, the number of haploid grains is generally less than 10%. High-speed and accurate identification of haploid grains can meet the!needs of engineering breeding. However, molecular biology and morphological identification methods commonly used in practical work are time-consuming, costly and destroying samples. It is difficult to select Maize Haploid grains efficiently and accurately. Relevant studies have proved that there are obviousoil content differences between haploid and diploid of high-oil maize. At present, low-field nuclear magnetic resonance technology can be used to detect oil content of maize and identify haploid according to its oil content. However, nuclear magnetic resonance (NMR) instrument has some weaknesses, such as high price, difficult maintenance, slow speed and low efficiency. It takes 4 seconds for each single-grain sorting. It cannot meet the needs of large number identification for engineering breeding. Using VIAVI near infrared spectrometer (NIRS) can achieve the detection speed of 0.25 seconds for each maize. The NIR technology is faster, cheaper and easier to maintain. The NIR identification method can replace the method of NMR. Qualitative identification of haploid by NIRS has achieved some results, but currently there are relatively few maize varieties collected in the study. The study only establishes models for haploid of one variety, and classifies haploid of that variety. There are no studies on identification of multiple hybrid haploids at home and abroad, but engineering breeding urgently needs a method to identify multiple varieties of maize haploids. In this paper, a method for identifying haploids based on deep belief network is proposed. DBN is a multi-layer deep neural network. Each layer is composed of a restricted Boltzmann mechanism. By using layer-by-layer training strategy, the problem that traditional neural network training methods are not suitable for multi-layer training can be solved. The comparative experimental results show that the identification model of multiple varietieshaploid established by DBN method has high classification performance and can meet the requirements of maize engineering breeding accuracy.