In-field pine seedling counting using end-to-end deep learning for inventory management

Abstract

<b><sc>Abstract.</sc></b> It is estimated that the southern U.S. produces over 1 billion pine seedlings for market sales per year with pricing varying from 50 to 350 dollars per thousand units. Accurate inventory of seedlings provides nursery management with 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 plant density in the field. A system with three cameras was developed to collect video from different drill rows on the seedling bed. The videos were preprocessed to restrict the region of interest to the center portion of the image in each camera and separated each drill row in individual videos. Two different input modalities of video and optical flow were evaluated as inputs. The inputs were fed to a convolutional neural network and long-term recurrent network to model the sequence of frames and regress to the seedling count for each plot. The mean absolute percentage error of our best performing model was 7.53% which is an improvement over the baseline manual sampling-based approach with 11.07%. The results showed that the proposed approach was able to count the seedlings in a crowded scene where multiple seedlings overlap their needles and stems in a complex outdoor and field conditions with higher accuracy than the current manual sampling-based approach. Therefore, the proposed system and results demonstrated the potential to replace manual counting and even provide further information such as a seedling density map over the field such that the nursery managers can plan better their practices.