Individual Assessment of Plant Growth Parameters Utilizing a Multi-Net Approach

With increasing demand for crop yields and increasingly harsh growing conditions due to climate change, breeding more resilient plant phenotypes becomes necessary. Assessment of plant growth while breeding is still primarily conducted manually on randomly sampled plants in the field. Automatic imaging systems such as UAVs are becoming increasingly common, making it possible to provide more complete datasets of entire fields or experimental parcels. To process those increasing volumes of raw image data, automatic processing from field-level data to plant-level growth parameters is required. Existing approaches in literature mainly focus on specialized assessment of specific plant phenotypes, growth stages, and assessment parameters, which provide limited scalability towards new crops or parameters. We propose to learn generalized features using a Multi-Net architecture with one single backbone shared by multiple heads, providing assessment for a selection of different plant parameters. We implemented a ResNet-based backbone in combination with Faster RCNN, Mask RCNN, and HTC assessment head architectures. We evaluated our approach both on experimental parcel data as well as the CVPPP dataset, which contains arabidopsis and tobacco plants of varying growth stages. ... mehrWe compare our Multi-Task Learning (MTL) model with its single task instances and all task combinations as permutations to showcase performance across multiple MTL scenarios. Our proposed model architecture allows for multi-plant and multi-parameter assessment, improving the plant selection process in breeding new plant phenotypes by enabling the evaluation of complete field datasets.