TC AgRA – Webinar 43 – Automated Weed Detection in Crops Using Flying Robots

Dr. Yasir Niaz Khan obtained his Ph.D. at the University of Tubingen in 2013. During his Ph.D. he conducted research on detection of terrain (ground surfaces) using a camera mounted on a flying and a ground robot. Upon completing his graduate studies, Dr. Khan started teaching robotics at FAST-NU, Lahore, Pakistan. He started a new robotics Lab at FAST-NU for robotics students to promote robotics in Pakistan. Dr. Khan supervised many national and international level robotic events held at FAST-NU where professors and students from different universities presented their works in robotics field. Dr. Khan joined the University of Central Punjab in 2015 and has started to extend his work further since he joined. He is now Director of Center of Robotics and Security that he established at the university. He is also running a project on Agricultural Automation for the last 3 years funded by DAAD Germany. The project includes visits to and from Germany of Pakistani and German researchers every year.

Abstract:
A farmer’s main goal is to produce a healthy and profitable crop. Sometimes, however, this goal is not achieved due to factors that affect plant growth, weeds being one of them. A crop badly infested with weeds normally fails in being profitable. Every year there is a huge loss of food crops like wheat, rice, maize, etc. due to weeds while the quality of produce from weed-infested fields is also inferior. Weeds affect not only the yield and quality of a crop but also raise the cost of production as weed control requires costly methods like labor force or uniform spraying. Automatic weed detection can help in the detection of localized weed infestation. Targeted spraying can then be applied according to the infestation rate, thus improving the crop yield per acre while lowering the cost of production and reducing environmental pollution. We use different image processing techniques for weed detection in different crops from a flying robot. Clustering and other machine learning and data mining techniques are applied to classify the crop video from a quadcopter based on weed density. Methodologies include classification and clustering techniques based on size features, FFT coefficients, GLCM, and Wavelet coefficients. They also involve transformations applied to local neighborhood. Most of the techniques applied to one crop are not generic enough to be applied to other crops. Different types of pre-processing and post-processing steps are applied to images depending on the quality of results and captured images like median filtering contrast enhancement, removing pixels less than or greater than some size etc. Accuracies achieved range between 66%-95% for different set of crops.