Information shared through social media and the internet may also be viewed as more relevant or persuasive to youth, with the social endorsement by trusted celebrities or peers . This is concerning, as an increase in social media use and novel potential for social engagement and peer network integration could increase youth vulnerability to cannabis marketing through social medical channels . Youth exposure to online cannabis marketing is especially concerning when it is accompanied by dispensary practices facilitating easy access to cannabis . Altogether, if replicated in a larger, more representative sample and during less unusual times, data showing that most exposures occur through online formats may suggest the need to better describe and reinforce online cannabis-related marketing to mitigate harms to youth. We also found that the timing and social context for cannabis-related marketing exposures occurred consistently throughout the week, mostly in the afternoon and the evening, while youth were alone and at home. This finding makes sense given that the majority of exposures occurred through the internet or public figures while youth were browsing social media online. As the majority of exposures also occurred through the internet or public figures, it is possible that cannabis advertisements influenced adolescents’ view of injunctive norms by suggesting high levels of peer approval of cannabis use and/or demonstrating or reinforcing positive outcome expectations related to cannabis use; this is particularly alarming because exposures generally occurred in the absence of family member who could buffer these effects . Research that identifies clusters in the context of cannabis-related marketing exposures is also important as it can inform cannabis marketing regulations, such as the need for tighter restrictions on marketing channels that frequently reach youth, especially in vulnerable contexts .
We also found that youth described cannabis marketing as relatively visually engaging. This is consistent with research demonstrating that companies marketing age-restricted substances create designs that likely appeal to youth, including bright colours,vertical grow rack cheerful messages, cartoon and/or animal characters, and other features explicitly prohibited by legislation . Unfortunately, we do not have sufficient data to conclude whether the vividness of ads impact adolescents’ cannabis expectancies or intentions to use cannabis. Policymakers and public health officials will likely benefit from larger and more detailed analyses of the features and content of cannabis ads that put them at the greatest risk for future cannabis use. Comparisons among cannabis naïve and experienced adolescents will likely be of considerable interest. Limitations of this feasibility study include its small sample size and its geographically unique sample of convenience. In Northern Ontario, adolescent cannabis use rates are high compared to major centres of the province . This smartphone owning sample may have been more willing or able to utilize EMA effectively than youth in the general population, although data now show that more than 85% of Canadian youth own and operate a smartphone . Second, information related to youths’ exposure to educational cannabis-related information and anti-cannabis information, and its subsequent effect on cannabis-related cognitions and cannabis use was not collected as part of this protocol, although it could be in the future. Lastly, this small sample size did not support meaningful statistical comparisons of random prompt and exposure occurrences, and assessment of differences according to demographic or SES factors. Larger studies, conducted beyond the immediate impacts of the COVID-19 pandemic, are needed to verify the type and impact on cognitions and cannabis use for cannabis-related marketing exposures. In conclusion, to our knowledge, we have provided the first example of an EMA protocol that adolescents can use to systematically demonstrate whether Canadian cannabis grow racks marketing regulatory efforts are comprehensive, effective, and the extent to which Canadian adolescents are exposed to cannabis marketing. Policymakers, educators, families and communities need to know the nature and extent of Canadian adolescents’ exposure to cannabis marketing and its impact on their attitudes, beliefs, and ultimately their decisions to use cannabis.
With data from larger, more diverse samples, this information could be used to hold companies accountable, to validate and enhance current regulations, and to minimize public harm of early cannabis use among youth. Land cover is one of the crucial indicators of changes occurring on the Earth’s surface . It is of great importance to study how land cover is changing over time to monitor climate change, food security, agricultural practices, crop yield prediction, and assess drought risk . Crop type classification using remote sensing is an active area of research . Among the applications of remote sensing in crop type classification is the use of image classification algorithms . In recent studies, a combination of several satellite datasets and various supervised classification methodologies have been used in crop mapping . Different types of machine learning algorithms exist. Among these algorithms are the Support Vector Machine , the Classification and Regression Tree , Random Forest, Artificial Neural Networks , K-Nearest Neighbour, Gradient Boosting Trees and Maximum Likelihood classifier. Several studies have used at least one of these classifiers for crop type identification . There is no consensus in the machine learning literature on the best-performing crop classifier. Recent studies have been focusing on comparing classifiers for crop identification through comparing their accuracy, performance, and computation time . In some cases, the SVM classifier was found to outperform ANN, RF, and maximum likelihood in terms of overall accuracy in classifying different crop types . Other studies have argued that SVM is more efficient at handling small data samples . In a comparison of SVM, normal Bayes, CART, and KNN, Qian et al. studied the effect of tuning parameters on classification accuracy using different training sample sizes. They reported that the SVM classifier’s accuracy is highly affected by tuning parameters, and SVM and NB classifiers were found to achieve higher classification accuracy compared to CART and KNN . Moreover, SVM and CART were found to be sensitive to changes in algorithm parameters, unlike RF which showed low sensitivity towards the tree and split parameters . Nonetheless, SVM and RF classifiers were found to be more robust compared to CART and performed better achieving higher overall accuracy . A cropland classification performed on different agro-ecological zones using RF, SVM, CART, and minimum distance classifiers showed that RF achieved the highest average overall accuracy and CART achieved the lowest overall accuracy. Comparison of the performance of SVM, RF, GBT, and Multilayer Perceptron to analyze the behaviour of different datasets in crop classification shows that GBT yielded better accuracy results compared to SVM and MLP .
Additionally, in a study that compares several classification algorithms for large area crop classification, RF outperformed other classifiers, followed by gradient boosting . RF, SVM and gradient boosting, being among the popular non-deep learning algorithms, were used as a reference by Zhong et al. , when developing a deep learning based model for crop classification. The deep learning-based model outperformed the non-deep learning algorithms, and among the non-deep learning classifiers, gradient boosting yielded the best results. The main challenge of land cover classification over large areas is the need to process “big data” and obtain cloud-free images. Traditional methods of processing “big data” are time-consuming, labor-intensive, and require significant storage capacity . Google Earth Engine can provide solutions when performing land cover classification. GEE is a cloud-based computing platform used for big data processing, it provides users with several datasets using a web-based Integrated Development Environment code editor without the need to download them, thus reducing storage space when processing data for large areas . In addition, built-in machine learning algorithms that simplify access to remote sensing tools are available on GEE which can be used in several studies including land cover and land use classification . Several studies have used GEE to perform land cover classification using several built-in classifiers . In a study that examined the efficiency of using machine learning algorithms in GEE to classify crops using multi-temporal satellite imagery from different sensors, the authors reported that the GEE platform enables better access to remote sensing data in terms of data processing . However, in terms of classification accuracy, the neural networks classification approach achieved better accuracy compared to GEE built-in SVM classifier . Fusion of optical and radar imagery in crop classification has been implemented in several studies . Optical imagery provides information on leaf pigments, water content, and plant health by using data from visible, near-infrared, and shortwave infrared portions of the electromagnetic spectrum . On the other hand, Synthetic-Aperture Radar provides information on canopy structure, surface roughness, soil moisture, and topography through the use of frequency and polarization . Denize et al. studied the benefits of using Sentinel-1 in winter land cover classification. Authors concluded that a combination of Sentinel-1 and Sentinel-2 time-series achieved higher accuracy compared to the use of Sentinel-2 or Sentinel-1 separately . The increase in accuracy of image classification due to the fusion of RADAR and optical sensors was also reported by Van Tricht et al. , where the fusion of Sentinel-1 and Sentinel-2 outperformed results of using single-sensor classification. Cannabis is one of the most highly demanded illegal drugs worldwide .
Thus, detection of illegal cannabis fields has become more challenging . Efforts to detect cannabis using image classification, remote sensing data along machine-learning algorithms have been limited due to difficulty in knowing cannabis plantation locations. In the Bekaa valley, cannabis cultivation has been considered a source of power and wealth. It has been a form of resistance against the French mandate in northwest Bekaa, and a revolutionary act since the French mandate prohibited the cultivation of cannabis. Cannabis production was encouraged even after independence , since development and management projects in Lebanon focused on the capital and major cities, neglecting the Bekaa region. In the 21st century, illegal cannabis cultivation continues in villages in Bekaa because the region is marginalized economically, is underdeveloped, and lacks basic infrastructure and employment alternatives. Thus, farmers in the region depend on exporting cannabis to afford their basic needs . The production scale of cannabis in Lebanon places it alongside Morocco, Afghanistan, India, and Pakistan as the top five countries for cannabis resin production in the world . Due to its high return value compared to other crops, cannabis is a major source of income for many farmers in the northern Bekaa valley. Many farmers continue to grow cannabis illegally in Lebanon given its minimal input requirements and much higher profit compared to other crops . According to local reports, it is estimated that one ha of cannabis can generate returns equal to ten times its establishment cost. The illegitimate nature of the industry has undermined the efforts to estimate the extent of its production and consequently its economic return, and it has posed limitations on traditional surveying as a feasible solution to scale and account for cannabis cultivation . Several cases have shown the inapplicability of site mapping due to the protective measures farmers use in order to hide their crops. Farmers tend to cultivate cannabis in inaccessible fields that are surrounded by tall crops or fencing walls. In addition to that, farmers were ready to engage in violent confrontation with the state security forces for protecting their plantations. For example, in 2010 cannabis farmers blocked the roads to prevent the national security forces from reaching their fields during a government-led eradication campaign . In this difficult context, remote sensing approaches appear to be the most promising and safest method for producing a reliable estimate of the cropped area and determining the exact locations of cannabis fields. Spectral reflectance difference between cannabis leaves and other leaves in the range 550–750 nm makes it possible to identify cannabis fields . Hyperspectral imagery and spectro-radiometry have been successfully used to detect cannabis ; ; ). Thiessen analyzed hyperspectral and multispectral imagery using machine learning techniques and standard image analysis to discriminate cannabis from other vegetation. Lisita et al. used high-resolution satellite imagery and object-oriented classification using SVM classifier to identify potential areas of cannabis cultivation. The classifier detected more than 50% of known cannabis sites. Ferreira et al. proposed a data-driven ensemble method to identify cannabis cultivation sites using remote sensing and deep learning techniques. The proposed models were successful in extracting features that can distinguish cannabis plantation sites. The effect of vegetation indices and textural features on detecting cannabis has not been studied yet. In this context, the capability of the Statistical Machine Intelligence and Learning Engine machine-learning classifiers that are now available in Google Earth Engine has not been fully explored. Although extensive research has been carried out on comparing classifiers and land cover classification, studies that discuss the use of classification algorithms and image fusion to identify cannabis cultivation areas are still limited.