טכניון מכון טכנולוגי לישראל
הטכניון מכון טכנולוגי לישראל - בית הספר ללימודי מוסמכים  
Ph.D Thesis
Ph.D StudentLati Ran
SubjectThree-Dimensional Driven Spatio-Temporal Model for Weed
Detection in Row Crops
DepartmentDepartment of Civil and Environmental Engineering
Supervisors Professor Sagi Filin
Dr. Hanan Isenberg
Full Thesis textFull thesis text - English Version


Abstract

Herbicides set the basis for most conventional weed management approaches. Although their usage has proven effective, increased public concern for environmental matters has put much pressure on farmers to reduce their use. In this context, site specific weed management can decrease herbicide usage by targeting only weedy areas.  However, implementation of such methods requires a means for autonomous weed identification. 3D models can provide weeds identification by estimating their growth parameter. Currently, only a few models have been proposed, where outdoor illumination is the main limiting factor and application is characterized by high-level of specificity. Additionally, no attempt whatsoever has been made to correlate biomass and the phenological growth phases of crop or weeds. Optimization in weeds control timing and rate can yield further reduction in herbicide usage. However, advanced knowledge of the weed biology and growth behavior under varying environments is required. Thus far, studies aiming to reduce herbicide usage have focused on spatial or biological aspects separately, while the relation between the disciplines has been ignored. This study develops a spatio-temporal model to characterize the relation between environmental conditions, plant biomass and its spatial growth parameters. Characterization is obtained via two sub models: 3D stereoscopic model for plant spatial parameters estimation, and environmental-related model for plant growth prediction. A 3D plant-oriented algorithm has been developed for estimation of spatial parameters. To handle the problematic matter of varying illuminations, a hue-invariant transformation was used and provided robust plant segmentation results under actual field conditions. The pixel matching part of the model integrated local and global optimization criteria. It was tested on different crop and weed species with varying canopy geometries, growth patterns and on different phenological phases. Application of these models provided accuracy, with a strong relation between the estimated and measured growth parameters (R2> 0.86 for all estimations). Moreover, a relation between plant volume and biomass was observed and characterized, turning the model into an effective tool for autonomous plant characterization and identification in versatile conditions. For plant growth evaluation, a photo-thermal model which integrates temperature and radiation measures for leaf area and biomass prediction has been developed, validated, and proven accurate. This model was then used for setting optimal control timing with lower rate. By combining biological and spatial aspects, results presented here can be a basis for future site specific and rational weed managements, and therefore assist in reduction of herbicide use.