WIRELESS SENSOR NETWORK FRAMEWORK FOR IMPROVING RICE PRODUCTIVITY IN NIGERIA

Abstract

Rice crop has been identified as a stable food for a large percentage of the world population and Nigeria, a country with over 170 million populations is not left out. Although the country spends huge amount of money to import this food crop due to its inability to produce enough for its populace as a result of some identified factors among which is low yield. Various parameters responsible for this low yield include climatic condition, soil condition, pest and diseases. Researches shows that pest is the most notorious biotic condition responsible for the low yield with over 15% of the cultivated crops being lost to these pests. It was also noted that Quealea birds are the major pest that destroy these grains. A solution through the use of sensors and machine learning technology was deployed to solve the stated problem. The research developed a wireless sensor network framework for improving rice crop productivity. The approach employed involves training machine to be able to detect, track and control actuators which will scare the birds away from cultivated grains. Images of the birds were captured by cameras. Three types of features: Haar-like, local binary pattern and histogram of gradient were extracted from images and compared to determine the most suitable feature technique for detection and tracking. The study shows that Haar-like feature performs better with accuracy of 72.5% and 75% precision while Local Binary Pattern (LBP) has a better sensitivity of 90%. Histogram of Gradient (HOG) has the worst performance having 50% sensitivity, 50% specificity, and 50% accuracy. The study also implemented and compared some tracking algorithm which are on-line boosting, multiple instance learning (MIL), kernelized correlation filters (KCF), tracking-learning detection (TLD) and median flow. Tracking-learning detection algorithm was identified to have the worst frame rate of 0.832 per sec but the best tracking failure 6 rate of 0.16%. On the other hand, kernelized correlation filters has the best frame rate of 137.0545 frames per seconds and worst tracking failure rate of 84.99%. In an effort to strike a balance speed and accuracy, an on-line boosting algorithm was chosen as the suitable tracking algorithm with an average tracking failure rate of 6.79% and average frame rate of 23.09 frames per seconds. The overall system performance was further evaluated using success rate. Out of 104 frames considered, 53 frames have an overlap ratio greater than 0.5 giving the detector success rate to be 51%. Information such as detection, tracking, centroid (x,y) and pixel area obtained were used to activate and control actuators such as light, bio-acoustic sounds and location of unmanned autonomous vehicle (UAV) in space in order to follow and chase the detected birds on the field. The system was developed and deployed on a rice farm to monitor parameters such as atmospheric temperature, relative humidity, soil temperature, moisture and bird pest. The system was able to report these sensed parameters to the cloud as well as prevent bird from invasion of the grains fields thereby improving crop yield.