We have proven the applicability of hyperspectral measurement combined with machine learning algorithms to predict the sex of C. sativa L. plants of cultivar Ferimon 12 from leaf reflectance spectra. Clearly, the biochemical information on the leaf surface and in the upper leaf cell layers at the early flowering stage is distinct enough to allow for the generation of mathematical models for the classification of male and female plants (Table 4). This assumption is supported by a study reporting pronounced differences in the content of polyphenols, flavones, and soluble protein, as well as differential activities of peroxidases and catalases, between leaves of male and female C. sativa L. plants (Elena et al., 2002). Results from this study matched the early observation of Talley (1934) that carbon to nitrogen ratios (C:N) of the above ground plant material differ between the sexual phenotypes, with higher C‐content in male and higher N‐content in female plants. Among the numerous reported phytocannabinoids, Δ9‐THC is generally considered to be more abundant in female plants, however being most concentrated in trichomes of the female inflorescence (leaves and buds); see ElSohly et al. (2017) and references therein. Also, several plant growth regulators (e.g., gibberellic acid, abscisic acid, and indole acetic acid) have been reported to be involved in the control of flowering and the sexual phenotype in hemp (Galoch, 1978; Hall et al., 2012). However, comprehensive profiling studies comparing male and female metabolic phenotypes are missing. To support the early prediction of sex in cannabis, kinetic studies uncovering the turning point at which the male and female phenotype can be distinguished metabolically would be needed as discussed below.

There is clearly large variability of chemical and morphological phenotypes of C. sativa L. (Grassi & McPartland, 2017; Petit et al., 2020; Strzelczyk et al., 2021), which in most cases determine the end use of the cultivar (Fike, 2016; Rehman et al., 2021). In our approach a panel of eight industrial hemp cultivars could be classified with high accuracy from leaf reflectance spectra (Table 5). This is in accordance with the results presented by (Lu et al., 2021) where machine learning based on regularized linear discriminant analysis achieved an accuracy of up to 99.6% in differentiating five industrial hemp cultivars. We assume that this high prediction accuracy mainly relates to distinct chemotypes of the cultivars in our study, which have been selected based on commercial relevance for hemp seed production and from different vendors (Table 1). As the geographical and ecological range of cannabis is unusually broad, substantial differences in the genotypes and thus in the chemotypes can be expected, as reported earlier by Lynch et al. (2016) and Borroto Fernandez et al. (2020) for European hemp varieties. This variability evidently generates a high demand for innovative technologies to cheaply, quickly, and non‐invasively determine cultivar identity in the growing cannabis market, for which the approach presented here proves to be ideally suited. However, despite the increasing scientific evidence for the applicability of spectroscopic approaches to discriminate plant varieties, such as shown for tea (Li & He, 2008), tomato (Xu et al., 2009), eucalyptus (Kumar et al., 2010), tobacco (Seiffert et al., 2010), grapevine (Diago et al., 2013), and hemp (Lu et al., 2021) their industrial applications are still missing (Dos Santos et al., 2013; Lopes & Sousa, 2018). Recently, an automated system combining hyperspectral imaging and machine learning for the classification of grapevine varieties under field conditions has been described (Gutiérrez et al., 2018).

Hemp is an oilseed crop, taking up more nitrogen and potassium, and similar amounts of phosphate as canola (Brassica napus L.) and performing well on highly fertile soil. Although it can be cultivated on a wide range of soil types, soil characteristics such as salinity, compaction and high acidity or alkalinity should be avoided (Kostuik & Williams, 2019). As the nutritional status of the plants determines the rate of seed set, good fertilization management is mandatory to maintain hemp yield in the field. In our study, a clear influence of soil management was observed, resulting in pronounced physiological differences between the plants grown on fertilized versus control soil. Leaf reflectance spectra were obviously distinct between the two groups, leading to a very precise classification for all cultivars investigated (Table 3; Table S1). We assume this to be related to different biochemical profiles in all cultivars related to the metabolized compost (better availability of nitrogen) in the plants on fertilized soil. Accordingly, increasing rates of animal manure, nitrogen, and phosphate fertilizers have been reported to enhance stem height and diameter, leaf, and stem weights, as well as the percentage of soluble extract (Laleh et al., 2017; Van der Werf, 1991).

The differences in soil fertility and therefore nutrition status of the plants have also influenced the early classification of sex from leaf reflectance spectra of young C. sativa L. plants (Figure 5). Prediction accuracies were highest for plants grown on fertilized soil in a cultivar‐dependent manner. However, improved prediction accuracies were not observed to be directly linked with accelerated plant development and thus altered flowering time (Figure 6; Table 7). Therefore, we concluded a combination of biochemical and morphological factors to be related to these differences, with likely higher impact of the chemotype. C. sativa L. is a complex species with overall highly variable morpho‐anatomical features, whereas most studies only concentrate on stem/fiber or trichome type and density characteristics (Raman et al., 2017). A study investigating diversity patterns among a collection of C. sativa L. genotypes identified the contents of bark fiber and cannabinoids as the highest discriminating factors (De Meijer & Keizer, 1996). Recently, substantial spatial gradients in secondary metabolite profiles in planta were observed which hint at organ and location‐specific regulation of accumulation in C. sativa L. (Bernstein et al., 2019), and which possibly vary between different genotypes. In addition, it was shown that C. sativa L. glandular trichomes, which are highly enriched on floral organs, alter morphology and metabolite content during flower maturation (Livingston et al., 2020), and similar changes may occur on leaf surfaces. We concluded that growing C. sativa L. on fertilized soil may have resulted in increased metabolite content and likely changed metabolite profiles in a cultivar‐specific manner, thus resulting in the differently enhanced prediction accuracies in our study. The observed reduced influence of flowering time on early sex prediction may relate to the time point of data acquisition, which we performed when the plants of all cultivars had developed between four and six leaf pairs. This assumption is supported by a molecular‐morphological study, in which microscopic analysis of male and female apices revealed that their reproductive commitment likely occurs as soon as the leaves of the fourth node emerge (Moliterni et al., 2004).

The most momentous effect on accuracy of early sex prediction was linked to the number of samples in the training datasets for the respective classes (male and female, see Table S2), with slightly higher sample numbers leading to significantly better accuracies. Class imbalance of training data, for example, one class heavily outnumbering the examples in the other class such as commonly happens in biological and medical datasets, has been frequently reported as a factor that negatively influences the performance achieved by existing learning systems (Batista et al., 2004; Vabalas et al., 2019). This situation typically leads to difficulties for the learning system to learn the concept related to the minority class. To overcome this problem the application of over‐sampling methods, for example, Random over‐sampling (Batista et al., 2004; Xiaolong et al., 2019) and the design of robust testing methodologies, for example, nested cross‐validation and train/test split approaches (Musto et al., 2021; Vabalas et al., 2019) have been suggested. Such approaches may also improve future prediction accuracies for the application presented here for cannabis sex prediction.

With respect to the limitations mentioned above, future improvements of the described approach for early sex prediction should include the systematic assessment and evaluation of data from cannabis plants during the course of development, from controlled environments and field trials, including larger numbers of cultivars, particularly medicinal types, and with a clear focus on sufficiently sized and balanced training datasets. In addition, limits for prediction accuracies need to be defined by C. sativa L. breeders and farmers, which allow for the economically viable application of spectroscopy‐based approaches for early sex determination in field and greenhouse environments. The methodology described here may also be applicable to other dioecious species of industrial and medical importance, such as date palm, kiwifruit, pistachio, yam, and jojoba, for which identification of the sexual phenotype at the seedling stage is of great importance to breeders and farmers for crop improvement and productivity (Sarkar et al., 2017).

Regarding cannabis, the application of classification approaches based on hyperspectral measurements combined with machine learning algorithms may possibly be extended to the prediction of Δ9‐THC and CBD contents, allowing state of the art non‐destructive analysis of plant chemotypes (Lopes & Sousa, 2018; Manley, 2014; Sanchez et al., 2020; Wang et al., 2016). Also, as shown for the classification of grapevine varieties from seeds (Zhao et al., 2018), the discrimination of cannabis cultivars may be possible even before sowing. For the monitoring and discrimination of species and cultivars in large and heterogenous field trials, airborne‐based approaches may be developed, such as shown earlier for cotton, rice, sugar cane, and chilies (Rao, 2008), and recently for hemp (Pereira et al., 2020). Generally, mathematical prediction models can be developed for a wide range of scales of spectral sensors, and thus may allow the creation of monitoring systems ranging from handheld to airborne devices in the future.