SMART WEED DETECTION AND ITS ACTIVE REMOVAL VIA VISUAL FEEDBACK SYSTEM
Advisor:
Dr. Muhammad Tahir
Abstract:
Weed control has been a troublesome issue since decades as they are the biggest enemy
of crops, causing the yield loss of nearly 20-40% annually. In Pakistan, the farming
communities mostly use cultural and chemical techniques, such as herbicides, for weed
removal which may harm the actual crop as well as pollute the environment. The
purpose of the project is to avoid these issues and develop an efficient automated system
that can safely suppress the unwanted plants. The project involves two phases; weed
identification and weed removal via closed loop control. Weed identification is done using
OpenCV and python. For training the system, machine learning approach is used. Data
set of tomato fields in initial growth stages is used. 250 image samples for plants and
weeds, each, are taken. The data set is pre-processed to remove the noise and enhance
the distinguishing features of plants. After this, geometrical, statistical, textural and
shape-based features are computed and each sample is labeled manually either as a
plant or weed. The labeled images are then used to train the SVM classifier model
based on RBF kernel for better detection. To optimize the classifier model, principal
component analysis is applied. The hardware mainly consists of a camera and a weeding
tool mounted on a robot. Camera captures images of the field, and based on trained
model, it detects weeds in the frame and find their x, y coordinates. The x-axis and z-
axis motion is managed by stepper motors while y-axis motion is achieved by actuating
gear motors attached with the wheels. Z-axis motion, covered by weeding tool actually
kills the weed. So far, 76% accuracy has been achieved on 100 test images. Running
time of the real-time detection algorithm is reduced by multi-threading.