Comparing pixel-based classifiers for detecting tobacco crops in north-west Pakistan

Pakistan faces heavy revenue losses in terms of one of its major cash crop i.e. Tobacco, due to the unavailability of accurate statistics of the total tobacco production. During the cropping season, there are many competing crops along with tobacco in the neighboring fields - making tobacco identification a challenging task. This study considers a pilot region of interest that spans over 64844 hectares, in the north-western Pakistan, covered through SPOT5 (2.5m) satellite imagery, acquired on June, 28, 2013. Two supervised pixel based classifiers: (1) minimum distance (MD) and (2) Spectral Angle Mapper (SAM) are compared and their overall accuracy discussed. The results show that there is no significant difference in the overall classification accuracy of MD and SAM. However, SAM performs better than MD with overall accuracy and Kappa coefficient of 76.56% and 0.7009 respectively. For the specific case of Tobacco crop, MD classifier has producer´s accuracy of 81.7% while SAM has that of 70.44%. The study also finds that Euclidean distance (in case of MD) and angle difference (in case of SAM) has no significant difference in classifying land cover types. It is also learnt that if area estimation is the objective, both of the classifiers will under-estimate tobacco covered area.