GLCM texture parameters such as contrast, homogeneity, energy, variance, and correlation have been proven to be successful in increasing classification accuracies by providing important textural characteristics that help in discriminating land cover . Additionally, GLCM texture metrics have been used in combination with SAR S1 data for crop discrimination . When classifying different crop types in an agricultural field, extracting information from neighboring pixels could be of great importance in improving classification accuracy. Texture features involve this information and help in identifying the intensity variations in an image which can further contribute to improving overall accuracy. We further observe that the addition of texture features can decrease the accuracy of a classifier, similar to what happened when using a CART classifier. Although RF and CART are tree-based machine learning algorithms, this decrease in accuracy was not observed using the RF classifier. This is because RF is an ensemble machine learning algorithm which makes it more robust and stable compared to a single decision tree classifier. In terms of comparing classifiers, the highest overall accuracy was observed using the SVM classifier, followed by GBT, RF, and finally CART. Similar findings were observed in a study that compares machine learning algorithms by Mustak et al. , where authors report SVM classifier achieving better accuracy in crop discrimination compared to RF and CART. Moreover, Sonobe et al. evaluated the potential of Terra-SAR-X data for crop mapping by comparing the performance of CART, RF, and SVM. Authors arrive at similar findings, with SVM being the optimum algorithm used. RF and GBT yielded similar results which have been observed earlier by Freeman et al. when predicting tree canopy cover in four study regions. We further assess the average accuracy in classifying cannabis by computing the F1-measure, which Yang et al. defined as the harmonic mean of the user’s accuracy and producer’s accuracy. In the 2016 classification, the average F1-measure achieved by RF, GBT, and SVM is 0.88. Thus, we assume that the three classifiers achieve similar results in identifying cannabis fields. In 2018, classification SVM and GBT achieve a similar average F1-measure , and RF achieves a slightly lower value . Since the classification models in 2016 achieved promising results in identifying cannabis, we use them to classify cannabis in 2017.
The models classify cannabis vertical farming with an average accuracy of 83% for SVM, 73% for RF, and 52% for GBT. Therefore, it is possible to detect illegal cannabis fields early in the season since it is usually harvested end of September – early October, and the classification period ends in August. The novelty of our study lies in classifying a non-dominant crop type with high accuracy as well as in the comparative analysis of the four most used machine-learning classifiers in Google Earth Engine. Previous crop classification research classifies crops that are dominantly present in study areas. To our knowledge, there has been no similar work carried out on cannabis classification. The Agricultural Improvement Act of 2018 permits the cultivation and legal trade of industrial hemp in the United States. This act defines hemp as Cannabis sativa and any part or derivative of the plant including seeds, extracts, cannabinoids, isomers, acids, salts, and salts of isomers with a total delta-9 tetrahydrocannabinol concentration below 0.3 % on a dry weight basis . This statute removed hemp-cannabis from its schedule I classification by using this definition to separate it from marijuana-type cannabis. Currently, there are no standardized methods to distinguish hemp from marijuana. Most forensic laboratories use chromatographic methods such as Gas Chromatography or High-Performance Liquid Chromatography methods coupled to mass spectrometry to quantitate the THC in suspicious plant materials. Furthermore, colorimetric tests that were once used to presumptively identify cannabis are not able to differentiate between hemp and marijuana, creating the need for an effective field test that can differentiate between hemp-type cannabis and marijuana-type cannabis. Hemp and marijuana are two different strains of the Cannabis plant with the main difference between the two being the concentration of cannabinoids contained in them. The two most important cannabinoids in these plants are THC and cannabidiol . THC is the cannabinoid that causes a psychoactive response in the body giving the person a “high”. It also has anti-inflammatory and analgesic properties, which make it desirable for medical use . CBD is also known for these beneficial properties but is non-psychoactive, so it does not give a person a “high” when used . Typically, hemp is CBD-rich containing low concentrations of THC causing its THC:CBD ratio to be below 1. Cannabis is considered marijuana when it has a concertation of total Δ9- THC ≥ 0.3% , but usually has a THC:CBD ratio above 1 . Elsohly et. al. reported that from 2009 to 2019 marijuana in the U.S. increased in THC potency across the decade from an average of 10% THC in 2009 to 14% THC in 2019. In 2019, the average CBD concentration in marijuana was found to be 0.6% , and that the THC:CBD ratio was above 20 across the decade . This difference in THC:CBD ratios can be used in the design of an effective field test for the identification of marijuana-type cannabis.
These tests are considered presumptive as they only indicate the possibility of the analyte being present in the substance . Until the Agricultural Improvement Act of 2018, the modified Duquenois-Levine test was the color test used to presumptively identify a suspicious plant material as cannabis. Although used for many years, the D-L test is known to produce false positives with reaction of molecules containing a resorcinol backbone and an aliphatic chain . Therefore, the D-L test is to produce false positive results from plants such as patchouli, spearmint, and eucalyptus. Furthermore, THC, CBD, and many other cannabinoids contain both a resorcinol group and an aliphatic chain, resulting in a D-L test that is not selective enough to differentiate between the cannabinoids. This shortcoming is the reason that the D-L test is no longer a suitable field test for the identification of marijuana-type cannabis. There is now an urgent need for color tests that can differentiate between hemp and marijuana . One colorimetric test that is currently being used to differentiate between hemp and marijuana is the 4-aminophenol test developed by the Swiss Forensic Institute in Zurich . A recent validation study has shown that a pink color forms when the THC:CBD ratio is below 0.3 and a blue color forms when the plant has a THC: CBD ratio above 3 . A confirmatory chemical test such as GC-FID or GC–MS is still required after a positive 4-AP test. The test requires the use of at least 1 mL of one of its reagents, 4-aminophenol, to produce a color result. Although the 4-AP test has demonstrated capability as a presumptive test for cannabis, it has also been reported that it may not be selective for THC. False positive results have been obtained with sage, oregano, and several cannabinoids, such as cannabinol.A more selective and smaller-scale alternative presumptive test could improve the presumptive confirmation for marijuana in the field. A colorimetric reagent that has been used for many years as a visualization reagent for cannabinoids when analyzing cannabis extracts through thin layer chromatography is the Fast Blue BB reagent . The FBBB test is selective among major cannabinoids, providing a red color for THC, an orange color for CBD, and a purple color for CBN. Ultraviolet-Visible Spectroscopy has shown that the FBBB + THC chromophore has an absorption band at 471 nm, which is responsible for its red color. The Almirall lab previously reported the structure of the FBBB + THC chromophore using results from high-resolution mass spectrometry and Hydrogen Nuclear Magnetic Resonance . It was determined that, in basic conditions THC becomes a phenolate anion and that this anion attacks the diazo group in FBBB at the para position to form the chromophore . A bathochromic shift results from the extended conjugation in the chromophore and the n→π* transition caused by the electrons in the diazo group of FBBB . In addition to characterizing the chromophore, the previous study evaluated the selectivity of FBBB for THC detection. Eight different types of tea, 3 hop products, and 3 authentic hemp buds were extracted and tested using FBBB. This test was performed by adding 10 μL of the extract to a filter paper, followed by 10 μL of 0.1% FBBB and 0.1 N NaOH. Extracts that were made from methylene chloride produced only 1 false positive with one of the teas . Of note, none of the hemp samples produced a false positive result, displaying an orange color indicative of CBD . These results support the selective nature of the FBBB test for use as a presumptive field test to distinguish between hemp, marijuana, and other plant materials. In the previous study, filter paper, a Capillary Microextraction of Volatiles device, and CMV strips were used as possible substrates to perform the FBBB test.
A CMV device is an open-ended 2 cm glass capillary tube that contains seven 2 cm by 2 mm glass filter strips have been coated with vinyl-terminated polydimethylsiloxane that was developed by the Almirall lab as an alternative to Solid Phase Microextraction. The modified glass filters that make up the CMV, known as Planar SPME , have excellent absorption/ adsorption capabilities and can withstand high temperatures. It was found that when the FBBB test is performed on one of the PSPME strips the LOD for THC was 100 ng, which is significantly lower than the known LOD for the D-L test, 5000 ng of THC . Using PSPME as a substrate is advantageous over regular filter paper since it can withstand high temperatures allowing the chromophores formed to be detected using DART-MS with very little background . In this current study, the capabilities of using FBBB as a presumptive field test to differentiate between hemp and marijuana are presented. We also report a fast and easy extraction method for plant material that can be used in the field. A previously reported substrate known as PSPME support was used for the FBBB reaction . Six cannabinoids, 5 retail hemp samples, vertical grow system 20 authentic cannabis samples, tobacco, hops, herbs, and essential oils were tested with the FBBB reagent. RGB numerical codes were obtained for each color result to confirm the color produced by the reaction in an objective manner. The fluorescence results of the FBBB + THC fluorophore is reported for the first time. The fluorescence spectra of the FBBB + THC product are distinguishable from the spectra of FBBB + CBD chromophores. The RGB score combined with the fluorescence of the FBBB + THC chromophore/fluorophore enhances the selectivity of the FBBB test for marijuana. Linear Discriminant Analysis was performed to determine whether FBBB could be used to classify cannabis correctly as hemp-type and marijuana type. The FBBB test was used to evaluate 6 different cannabinoids, 5 commercial hemp strains, 20 cannabis samples, and various herbs and spices. It was determined that when FBBB reacts with THC, it forms a red chromophore that fluoresces under 480 nm light. Conversely, when reacted with CBD or CBD-rich products, such as hemp, an orange chromophore is formed, and this chromophore does not fluoresce. This is the first time, to the author’s knowledge, that the fluorescence of the FBBB + THC chromophore/fluorophore is reported for a colorimetric test. This fluorescence is easily visualized using a portable Dino-Lite microscope and its spectra obtained with a VSC2000 spectrometer. The intensity and wavelength of the fluorescence for the chromophore combined with the distinct red color it displays makes for a more selective and sensitive test to differentiate between marijuana and hemp. The structure for FBBB + THC has been previously determined by the Almirall lab, as shown in Fig. 1 . The chromophore results from an extended conjugation of π-bonds decreasing the distance between energy transitions between the ground state and excited state. This extended conjugation causes a “red shift” of the FBBB chromophore, which is responsible for the red color and the fluorescence that is observed when THC reacts with FBBB. One theory for CBD + FBBB lacking fluorescence intensity is that CBD has a less rigid structure than THC. It is known that structure rigidity and a fused ring structure increases the quantum efficiency, and therefore fluorescence of a molecule. Since CBD is less rigid than THC and does not have a fused ring structure, it is prone to relaxation through internal conversion rather than through radiative means . Therefore, FBBB + CBD likely relaxes through nonradiative mechanisms, which decreases overall fluorescence. The difference in both color and fluorescence that is observed for FBBB + THC and FBBB + CBD is an advantage that the FBBB test has compared to other tests for presumptive analysis of cannabis, which only use color.