UC Davis

Agricultural food crop production must increase 100-110% by 2050 in order to meet the predicted global population and production demands. In order to meet the increasing food demands from the growing population, as a society we are looking for ways to increase food yield without needing to further exploit land and water resources. Plant breeders look to solve this problem by breeding novel high-yielding crop varieties that can produce quality fruit in extreme weather and limited water conditions. This includes crossbreeding commercially available cultivars with wild types shown to react favorably in times of drought and heat stress in nature. The two key components in breeding commercial crop cultivars are quantifying genotypic and phenotypic data.Prior to second generation sequencing, the bottleneck in producing commercial crop varieties was the inability to accurately and efficiently sequence a crop genome. The second generation of sequencing has simplified the process of genome mapping and allows for the identification of candidate genetic markers to utilize marker assisted selection (MAS). With these developments, the current bottleneck in producing new crop varieties is now phenotyping plant traits in in-field breeding trials. There is an increased need for higher throughput in relating genetic material to phenotypes. Current in-field phenotyping methods rely on humans to go into the breeding fields to detect these phenotypes. These current methods are time consuming, labor intensive, and subjective. These phenotyping trials rely on repetition of genetic material in a study to maximize time and space and reduce the amount of human labor is needed. In order to meet current and future food demands, we need to look at new methods of high-throughput phenotyping to help plant breeders create new crops more efficiently. In this study we look at three aspects of high-throughput phenotyping: developing and deploying a multi-camera and multi-sensor in-field phenotyping system for Solanaceae crops, developing algorithms to use with the phenotyping system to detect phenotypes in tomatoes, and exploring new sensor technology to develop low-cost methods of detecting green fruit. A high-throughput phenotyping (HTTP) system was developed to be a high clearance tractor fitted with 9 color DSLR cameras, 9 pseudo-NDVI cameras, 2 Time of Flight cameras, 2 temperature sensors, 8 custom LED modules, with infrared and broad-spectrum white LEDs, 1 RTK-GPS, 1 National Instruments High Speed Controller, 1 Dell desktop computer, and 1 National Instruments USB controller. The system was used in 2018 and 2019 to collect multi-sensor data in a processing pepper and tomato breeding trial in Davis, CA. Images from a single camera from the HTPP system was used to develop algorithms for object detection in tomatoes throughout the growing season. The algorithm utilized Mask R-CNN for object detection of flowers, green fruit, ripe fruit, and sunburnt fruit with an average mAP of 21.1% on a validation set. The number of ripe fruit and green fruit on the last date were used to predict the weight of ripe fruit and green fruit from manual phenotyping at harvest, the R values were 0.407 and 0.201 in 2018 and 0.202, and 0.142 in 2019 respectively. A pseudo-NDVI camera was used to test if adding non-visible wavelengths to imaging would impact the detection of green fruit. DSLR cameras were modified to allow infrared and near-infrared wavelengths to be imaged, in hopes that this would be a low-cost method for incorporating multi-spectral imaging in the HTPP system. An infrared camera and color camera were compared for green fruit detection in 2019. 85% of the plant-plots imaged had more green fruit detected in the color images than the infrared camera, which could imply double counting in the color images. More testing and ground truth data needs to be done to get conclusive answers about the green fruit detection in infrared. In conclusion, this dissertation focuses on developing high-throughput phenotyping methods that rely on 2D imaging of in-field plants. The ability to quickly process quantitative phenotypic data is important for the development of new crops as the current bottleneck in plant breeding is phenotyping. This dissertation shows that there are new and old technologies and techniques that can be applied to high-throughput phenotyping for in-field phenotyping.