Signal and image processing technology for smart agriculture applications

This thesis is concerned with development of signal and image processing technology for smart agriculture applications, with a particular focus on applications in automatic weeding systems. Developing an automatic weeding system requires robust detection of the exact location of the crop to be protected. Computer vision techniques can be an effective means of distinguishing crops from weeds and determining their locations. To achieve this, several practical issues need to be addressed such as weather variability, presence of shadows in sunny conditions, natural similarities between the target object (weed or crop) and the background, occlusion of objects of interest, and unexpected changes in camera parameters. This thesis addresses a number of these issues. Firstly, a comprehensive study was conducted on image-based plant segmentation techniques (identifying plant from a background of soil and other residues). Three primary approaches, namely, (i) colour index-based segmentation, (ii) threshold-based segmentation, (iii) learning-based segmentation are discussed. The challenges and some opportunities for future developments in this space are identified. Secondly, a novel algorithm based on colour features and shape analysis is proposed to detect cauliflowers on a frame-by-frame basis from video acquired under various weather conditions (cloudy, partially cloudy, and sunny). The algorithm was tested under different weather conditions and achieved a detection performance of 98.91% and precision of 99.04%. Then, in order to increase system robustness, the detection algorithm was extended through the addition of object tracking. A multi-object tracking algorithm based on Kalman filtering and the Hungarian algorithm was applied. With the help of the tracking algorithm,detection failures were reduced, especially in sunny conditions, such that overall detection performance was raised from 97.28 to 99.34%. Overlapping between plants (depending on the plant growth phase) is one of the most challenging problems that face computer vision techniques in real conditions, especially for plant segmentation and classification. In order to solve this issue, a novel approach based on main stem feature detection is proposed. Results of evaluation of the proposed algorithm show that the majority of plants were correctly detected with distance error of less than one centimetre, even in occluded conditions. Finally, a comparative study of plant classification using deep learning approaches and traditional approaches was conducted. Two well-known deep learning architectures (AlexNet and GoogleNet), and three based on Support Vector Machine (SVM) with different feature sets (Bag of Words in L*a*b colour space feature, Bag of Words in HSV colour space, Bag of Words of Speeded-up Robust Features (SURF)) were applied. Results show that the best overall performance was achieved by deep learning-based approaches (AlexNet and GoogleNet), while the SVM-based approaches achieved close to the same performance