Cannabis has attracted significant attention in recent years due to its medicinal and recreational properties. However, certain cannabis varieties contain high levels of tetrahydrocannabinol (THC), which can pose health risks. This underscores the need for reliable methods to detect and classify diferent cannabis varieties. Traditional manual classification and sorting techniques are often time-consuming and error-prone, highlighting the potential of artificial intelligence (AI) as an efective alternative. Despite its promise, the application of AI, especially deep learning, faces challenges such as limited availability of labeled data and the necessity to deploy models on mobile devices with constrained computational resources. To address these issues, this study explores the use of transfer learning for classifying cannabis seeds. Transfer learning mitigates data scarcity by relying on pre-trained models, while computational eficiency is tackled by selecting architectures optimized for mobile environments or characterized by relatively low resource demands. The experimental evaluation, involving a new dataset comprising two cannabis seed varieties, demonstrates that deep learning models employing transfer learning can achieve high classification performance, even under resource-limited conditions.
Cannabis is the botanical name of a genus within the Cannabaceae, the same plant family that contains
hops. The genus includes three species, Cannabis sativa, Cannabis ruderalis and Cannabis indica [
phytocannabinoids. Tetrahydrocannabinol (THC) is the most well-known cannabinoid. It
shows considerable medical benefits. Indeed, THC relieves symptoms of sleep disorders, anxiety, and
insomnia and acts as an antidepressant. However, high doses of THC can hurt thinking, concentration,
perception, and mental function, potentially leading to behavioral disorders, hallucinations, delusions,
or psychosis. Because of these side efects and risks, cannabis plants and relative seeds with high level
of THC are illegal in many countries [
Conventional manual classification and sorting procedures are time-consuming, labor-intensive,
and prone to human error. In some ways, DNA-based forensic botany techniques have outperformed
conventional chemical methods in the analysis of cannabis. Nevertheless, prior eforts to find and
confirm genetic markers on these synthase genes for crop type identification have run into problems [
In response to these challenges, artificial intelligence are emerging as a promising solution. In literature,
deep learning methods have been widely used for classifications of crop seeds [
CEUR Workshop
ISSN1613-0073
[
Unfortunately, one major challenge in deploying deep learning models, especially for specialized tasks such as identifying cannabis seed types is the limited availability of labeled data. Deep neural networks, particularly convolutional architectures like AlexNet, GoogLeNet, or RetinaNet, typically require large datasets to generalize well. Moreover, to make practical the use of these models, deep learning methods should run on mobile devices. However, mobile and embedded devices face hardware limitations that restrict the deployment of complex deep learning models such as limited memory and storage and low processing power. The combination of a small dataset and limited computational power complicates the use of conventional deep learning approaches on mobile platforms. Even if a model is trained ofline on a high-performance server, deploying it to a resource-constrained environment requires adaptation.
Starting from these considerations, in this work, a comparison of CNN architectures using transfer
learning has been carried out to classify cannabis seeds. Transfer Learning is one of the most widely
used pretraining techniques which requires first training a baseline model on a large dataset and
then fine-tuning it on a smaller target dataset. To create a solid foundation model, transfer learning
uses parameters already trained on other source data rather than explicitly training the model on a
comparatively small target dataset [
In our study, four diferent state-of-the-art pre-trained CNN architectures have been used to extract
features and classify images of cannabis seeds: MobileNet [
The MobileNet family comprises multiple variants of CNNs designed to be lightweight and eficient,
making them ideal for use on mobile and embedded devices with limited computational resources.
It breaks down the standard convolution into two separate operations, depth-wise convolution layers
and point-wise convolution layers, reducing the number of parameters and computations required
while maintaining or even improving accuracy. MobileNetV2 [
EficientNet is a family of CNN architectures developed by Google AI, designed to achieve high accuracy with significantly fewer parameters and lower computational cost compared to traditional models. The key innovation behind EficientNet is the use of a compound scaling method, which uniformly scales the network’s depth, width, and resolution using a set of fixed scaling coeficients. EficientNetB0, part of the EficientNet family, is the variant used in our experiments. This variant is specifically designed to achieve high performance while being computationally eficient. EficientNet-B0 starts with a stem convolution and is followed by seven stages of Mobile Inverted Bottleneck Convolution (MBConv) blocks, each with specific kernel sizes, strides, and expansion ratios. Inside the convolutional layers and MBConv blocks, the activation function used is Swish. Instead, at output layer, softmax function is typically used.
Xception is a deep convolutional neural network architecture that takes the original Inception idea to its logical extreme by completely replacing Inception modules with depthwise separable convolutions. In detail, Xception’s architecture is composed of three main flows: 1) Entry Flow composed of a few regular convolutions followed by depthwise separable convolutions; 2) Middle Flow composed of identical blocks of depthwise separable convolutions with residual connections inspired by the success in ResNet; 3) Exit Flow composed of a final set of depthwise separable convolutions and max-pooling, followed by global average pooling and fully connected layers. Its hierarchical structure aids in learning hierarchical representations and facilitates the flow of information through the network.
NasNetMobile is a convolutional neural network designed using Neural Architecture Search (NAS), i.e. an automated process that uses machine learning to discover high-performing architectures. Indeed, NAS framed the problem of finding the best CNN architecture as a Reinforcement Learning problem through its three main constituents: search space, search strategy, and performance estimation. The search space defines the set of possible architectural choices that can be explored. The search strategy determines how the NAS algorithm explores the search space to find the best architecture. Once a candidate architecture is selected, the evaluation strategy measures its performance.
This section is devoted to describing data, experimental setup and results of our comparison study.
A total of 84 cannabis seeds, including marijuana and hemp from various geographical origins, were used in this study. Seeds were legally purchased online or from authorized cannabis shops: • Italian hemp (N = 10), CB Weed (Le Marche); • Czech hemp (N = 10), Cannandorra Hemp (CZ-BIO-002); • Danish hemp (N = 10), Raab Vitalfood (DE-OKO-001); • Danish hemp (N = 10), Reformhaus Hemp (DE-OKO-003); • Spanish hemp (N = 10), Gramso (Comunidad Valenciana); • Marijuana (N = 34), Royal Queen Seeds.
Despite their diverse origins, all samples were included in a binary classification task (hemp vs. marijuana).
Each seed was imaged using a fully automated inverted fluorescence microscope (Leica DMI6000b). Images were captured at 4000 × 3000 pixels (96 dpi, both horizontal and vertical resolution) and saved in .JPG format.
Various preprocessing methods were used on our dataset to improve the quality of the input data, make
the model training more eficient, and ensure model generalization. These images were pre-processed
in the following sequence:
• Data labeling to categorize hemp and marijuana classes;
• Transform non-numerical labels into ordinal encoding scheme;
• Seed segmentation using Rembg [
Figs. 1 and 2 show the efect of the preprocessing operations on the images, after segmentation and resizing. From the images it is possible to deduce that the transformations have not worsened the quality of the images and the salient features are still well visible and optimized for the training process. (a) Original hemp seed (b) Hemp seed after preprocessing
After these initial image preprocessing steps, we split the dataset in the ratio of 80:20 for training and test folders, ensuring class balancing. Validation data have been obtained from training data with a ratio of 90:10 for training and validation folders for monitoring the model’s generalization performance. For feature extraction, we have used multiple pre-trained CNN architectures as mentioned before from Keras Applications. Default hyper-parameters were used. In particular, = 1 for MobileNetV2 allows a good trade-of between performance and computational cost.
We get pre-trained weights alongside each model. All of them were trained on the well-known
ImageNet dataset [
The added layers in our networks are the GlobalAveragePooling2D and Dense layers. The GlobalAveragePooling2D layer is used to extract spatial information from feature maps. It’s instrumental in reducing parameters and simplifying model architectures, as it contains fewer parameters than the Flatten layer, which reduces the risk of overfitting and helps build a more eficient model. The final layer of the model is the Dense layer, The goal is to use the pre-trained model, or a part of it, to pre-process images and get essential features and pass these features to this new classifier with no need to retrain the base model. The Dense layer contains 1 neuron because there are two possible classes, hemp and marijuana, and a Sigmoid activation function to output a probability between 0 and 1.
Details of the training settings are given in Table 1. In this configuration scenario, all 4 models contained frozen and trainable parameters whose size is displayed in Table 2
Performance evaluation methods such as Accuracy, Precision, Recall, and F-score were used to evaluate models created for our binary classification task. These metrics were obtained from the confusion matrices where true positives, false positives, true negatives and false negatives are reported. It is important to note that, in a forensic context, if the positive class is defined as representing illegal seeds, the presence of false negatives in cannabis seed classification means that some illegal seeds may go undetected, potentially leading to serious operational consequences.
Table 3 shows the training and test performance metrics of each compared model taken as a weighted average (taking the support of each class into account) on the acquired dataset.
(a) MobileNetV2 loss
(b) Xception loss (c) NasNetMobile loss (d) EficientNetB0 loss
EficientNetB0 produced worse outcomes on both training and testing data. Neither meaningful patterns nor efective generalization were learned by the model. Due to the small dataset size and lack of training time, the model may not have learned the relevant patterns in our dataset, which could lead to poor performance. As shown in Fig. 3, validation loss remains high after 20 epochs, suggesting the need to use more epochs and data-augmentation technique to fine-tune the weights properly.
MobileNetV2, Xception and NasNetMobile models confirmed their strength even in the presence of small data, and they are able to converge, overcoming the trade-of between the amount of data vs. model complexity. As expressed in Table 3, MobileNetV2 and Xception achieved the best metrics for marijuana/hemp classification, achieving a 100% score for each metric performance on test data. They are (a) MobileNetV2 results
(b) Xception results (c) NasNetMobile results (d) EficientNetB0 results followed by NasNetMobile, with lower performance in terms of accuracy, Recall, F1-score, and Precision (82% for all metrics). The above metrics are calculated with the help of confusion matrices reported in Fig. 4 which provide a detailed breakdown of how well each classification model performs. Confusion matrices confirm the robustness of the models, with the exception of the EficientNet architecture which, as can be seen, incorrectly classifies all marijuana samples as hemp seeds.
To conclude, considering also the number of trainable parameters in the comparison, the MobileNetV2 model emerges as the most eficient choice, as it achieves the same performance of Xception while requiring significantly fewer parameters to be trained.
As part of our study, four state-of-the-art CNN-based transfer learning models, such as MobileNetV2, Xception, EficientNetB0 and NasNetMobile were trained on a new dataset of marijuana and hemp seeds. All four architectures are designed with eficiency in mind, specifically focusing on achieving high accuracy while keeping the model size and computational cost low. The best accuracy results were obtained with MobileNetV2 and Xception architectures, reaching 100% accuracy on test data, correctly classifying each seed class. However, when considering the number of trainable parameters, MobileNetV2 proves to be the best option, matching Xception’s performance with a much smaller model size. In general, except for EficientNet, the considered CNN architectures using transfer learning ofer good performance and computational eficiency, making them a good choice for mobile applications also in the forensic context.
This study was funded by Ministero dell’Università e della Ricerca (MUR) of Italy in the context of project denoted as BLOODSTAIN in the program PRIN 2022 (grant number E53D23008040001).
The authors have not employed any Generative AI tools. M. Bernstein, et al., Imagenet large scale visual recognition challenge, International journal of computer vision 115 (2015) 211–252.