Estimation of Cotton Population Density in Field Conditions Using Deep Convolutional Neural Networks

Population density is a major component of cotton fiber yield. It can be estimated by combining laboratory test of seed viability and sub-sampling of field germination, both of which are subjected to errors. The overall goal of this study was to develop an imaging-based method to count germinated plants in field conditions using convolutional neural networks. An unmanned aerial vehicle (UAV) with a consumer-grade color camera was used to image a cotton field containing 500 plots on the 22^nd day after planting (DAP 22). Collected color images were used to reconstruct an RGB-D image of the entire field through commercial software. In the RGB-D image, individual cotton plots were segregated based on GPS information. Within each plot, single-plant images were obtained using color and size filters. Each single-plant image was rotated 360 degrees with an interval of 90 degrees to increase the number of single-plant samples. In addition, 20 sub-images were randomly selected to generate non-plant images. A 20-layer convolutional neural network (CNN) with residual architecture was trained using the selected single- and non-plant sample images. The trained CNN was applied to RGB-D images of individual plots to identify and count each single-plant. Experimental results showed a high correlation between imaging-based and manual counts, indicating a high accuracy of the imaging-based method. Therefore, the developed counting approach could be particularly useful for calculating population density and thus for predicting cotton fiber yield.