A new remote sensing-based index for discrimination of Cannabis sativa

Abstract

Cannabis Sativa is an annual crop that is cultivated in the Bekaa valley of Lebanon. Cannabis is an important crop for farmers in north Bekaa due to its low production cost and high reutrn compared to other crops. Since cultivation of Cannabis is prohibited, no statistics are available concerning its cultivation areas and annual production. In this research, remote sensing techniques were used to detect the Cannabis Sativa fields in the Bekaa valley. Different satellite imagery were utilized in the study including Worldview-3, Sentinel-2, RapidEye, Landsat 7, Landsat 8, and ASTER. The research was developed by digitizing surveyed fields of the aforementioned crops in GIS, updating vegetative status of all fields to match with the date of the satellite imagery, and discriminating Cannabis Sativa fields by image classification techniques. Two pixel-based image classification techniques were used for Cannabis detection: maximum likelihood classifier (MLC) and decision tree (DT). MLC is a traditional classifier that is based on training data and statistical parameters, whereas DT is a conditional classifier that is independent of statistical assumptions. Both classification methods were evaluated by accuracy assessment using digitized data not used for training. Classifications were evaluated by analyzing user’s accuracy, producer’s accuracy, and total accuracy. MLC was applied to all satellites and proved the ability of Landsat 7 and 8, Sentinel-2, and RapidEye to discriminate Cannabis. DT had several advantages over the MLC by not requiring training data and giving higher accuracy results. Results show that it is best to implement the developed DT with Landsat 8 imagery between May and June at CDI threshold of 4.8 and FOV of 0.36 to detect Cannabis Sativa fields. Using the DT classification method, the total cultivated areas of Cannabis in the northern-central Bekaa were estimated to be in the range of 1,500 - 2,200 ha for year 2015 and 2,700 - 4,000 ha for year 2016. The research results are promising f