Precision Agriculture Systems for the Southeast US Using Computer Vision and Deep Learning
Type of DegreePhD Dissertation
MetadataShow full item record
Recent advancements in sensors, machine vision and deep learning with the development of efficient algorithms has enabled new opportunities in precision agriculture. In this dissertation three different projects are presented. First, an automated pine seedling counting for nursery management. In nursery management, accurate inventory of seedlings provides insights into how many seedlings can be sold and/or if there is any loss due to washout, mechanical damage or pest/diseases that can still be mitigated. In this study we developed a system to count pine seedlings in production sites and map the seedling density in the field. The mean absolute percentage error (MAPE) of our best performing model was 7.53%, which is an improvement over the baseline manual sampling-based approach with a MAPE of 11.07%. The results showed that the proposed approach was able to count seedlings in a crowded scene under complex field conditions with higher accuracy than the standard manual practice. Second, an mmWave radar-based peanut yield monitor, an essential equipment for precision agriculture is presented. During the harvest of peanuts, it is common for foreign materials to run through the pneumatic conveyor in the harvester, making the environment inside the pneumatic conveyor harmful for components that are fragile. Therefore, a millimeter-wave FMCW radar-based mass-flow sensor was developed to monitor yield during peanut harvest. After evaluation of 5-fold cross validation, the best performing models achieved an RMSE of 0.14 kg/s (19 lb/min) and a sMAPE of 15% while having an R2 of 0.85 for the research-scale combine and an RMSE of 0.52 kg/s (69 lb/min) and a sMAPE of 10% while having an R2 of 0.71 for the commercial-scale combine. Though the results are promising, further field investigation is necessary to evaluate the effects of different field conditions. Finally, a broiler activity index measurement is demonstrated using a detect-and-track pipeline and a age-compensated activity index. Precision management in poultry farming, utilizing ad- vanced technologies like computer vision and lighting control, has emerged as a promising approach to enhance productivity, meat quality, and animal welfare. This work generated multi-animal detection-and-tracking dataset from top-view videos of broilers at different growth stages and dif- ferent times of a day. Accuracy and efficiency of different state-of-the-art object detection (YOLOv8 and YOLO-NAS) and trackers (SORT and ByteTrack) were evaluated and compared regarding the generation of a new broiler activity index. The fastest pipeline (YOLOv8n+SORT) achieved 123 frames per second with a correlation of 96.4% with the proposed broiler activity index. Once the behaviors are characterized the pipeline can be used to generate feedback signal for control of the broiler houses, providing better animal welfare.