Investigating Relation between Open Boll Count and Cotton Seedlings Emergence Using Unmanned Aerial System Images and Yolo, a Deep Learning Framework

Accurate estimation of cotton yield at early growth stages can make agricultural resource management more efficient throughout farming season. In this study, we try to estimate open boll count from number of cotton sprouts using unmanned aerial system (UAS) images and YOLO (You Only Look Once), a deep learning framework for real-time object detection. YOLO uses single convolution neural network to predict the boundaries of existing objects and class probabilities simultaneously therefore it is suitable for large dataset as UAS images. The image data of this study were acquired at Texas A&M AgriLife Research and Extension Center at Corpus Christi, Texas, in April of 2017 when number of days after planting was less than 20. Total 466 images of 20 Megapixels were taken for 80 m by 170 m area where spatial resolution per pixel was about 0.55 cm. A YOLO network was trained from sub-dataset from original UAS images, and trained network model was used to locate center of each cotton sprout. Count of cotton sprouts and open bolls were aggregated using square grid along the rows of cotton plots. The correlation between number of cotton sprout and open boll is calculated and presented per each variety used in this trial.