Identification and Counting of Sorghum Panicles Using Artificial Intelligence Based Drone Field Phenotyping

Abstract

One of the most promising and difficult challenges for field phenotyping is accurate and reliable counting of sorghum panicles using drone imagery both from RGB and multispectral cameras. In this paper, we present a hybrid Machine Learning method for sorghum panicle identification and counting.The methodology first consists in building a Machine Learning classifier following the two most used methods in the literature for drone and agriculture applications: Support Vector Machine Learning (SVM) and, Artificial Neural Networks (ANN). The present dataset includes 5300 images, and 60% of the dataset were used for training and 20% for testing and validation. Following the results obtained from these models, image segmentation using super-pixel affinity propagation and k-means clustering was used based on simple linear iterative clustering. With an accuracy of 99%, SVM gave a superior performance also in terms of precision and kappa when compared to the ANN model whose accuracy was 98%. Concerning the SVM, a radial basis kernel was used, and the sigma parameter was kept constant at a value of 5.6 determined analytically.