Application of pre-trained deep learning models for sorghum panicle characterization

The capability to characterize various plant phenotypes from images is increasingly recognized as vital to crop improvement. In sorghum crop improvement, inflorescence (panicle) data play a vital role in the understanding of genetic diversity among varieties, yield estimation and aid the selection of new cultivars. The adoption of small unmanned aerial systems (UAS) is enabling frequent collection of the needed crop image data, but at the same time creating massive datasets which are difficult to analyze efficiently. The emerging field of deep learning promises unparalleled performance in detecting and segmenting image objects of various kinds from large datasets. Deep learning approaches are also easier to adopt due a wide number of free pre-trained models. This study evaluates a number of deep learning models (ResNets by Microsoft Research, GoogleNet by Google) for detection and segmentation of individual sorghum panicles in UAS images collected over the sorghum trials at Texas A&M university farm in Burleston County TX for 2016 and 2017 growing season. We apply panicle detection and segmentation results to estimate panicle counts by plot and individual panicle measurements (panicle width and length). We compare the derived panicle counts and panicle sizes with field-based and manually digitized measurements in selected plots and study the strengths and limitations of each model for sorghum panicle characterization.