The Data and Decision Sciences Lab is constantly working with local community partners on several data related projects that have a local and global impact.
Crop Detection in Aerial Imagery Data Using Deep Learning
Community Partner: University of Nebraska–Lincoln – Department of Agronomy and Horticulture
Team Members Involved
The United States produces more corn than any other country in the world. The land used to plant corn is over nine million acres. The production of corn plays an indispensable role in the U.S. economy. Increasing corn production while reducing the amount of land used to grow corn have a positive effect on the U.S. economy. To achieve this goal, it is necessary to analyze the genotype and the biomass of corn. In order to analyze the biomass of corn, one must first understand the difference in genotypes as well as the order of the genotypes. A cornfield located in Lincoln, Nebraska, consisting of thousands of different genotypes of corn was chosen to study the genotype and corresponding phenotype. The dataset of this project was collected by using Drone technology. The data was provided by the Yang Lab. The purpose of this project was to use deep learning methods to classify the corn, weeds, and soil based on the Unmanned Aerial Vehicle (UAV) images. The dataset was obtained directly from the cornfield with the help of drones. Thus, each pixel in the image has no label. The resolution of each image is 5472 x 3648 pixels. This project manually cut and labeled 600 images from the original UAV images, with each image being 198 x 198 pixels. There were 200 images in each category. Furthermore, 498 images were used as training set and 102 images were used as validation set to train basic Convolutional Neural Network (CNN) model and more complex model – Inception V3 model. The training accuracy of the two models were around 97%. The original UAV images were divided into two different sizes of images, with one being 202 x 202 pixels and another size being 96 x 96 pixels. The project then used two CNN models to train two different sizes of images separately. The results showed that the model using small size images to test had better prediction.