Comparison of Deep Learning and Object-Based Image Classification Methods: Identification of Horses from RGB Imagery

Abstract

The paper describes the utilisation of remotely sensed RGB data to support routine monitoring of horses in a natural environment on demand. Data are sensed using an unmanned aerial vehicle (UAV). UAVs provide very high spatial resolution data sensed at a low altitude on demand. Sensing is limited by weather conditions and legal rules only. Terrain does not need to be accessible. The article provides a comparison of several classification methods, namely object-based classification methods and Deep Learning classification. Namely Maximum Likelihood, Random Trees, Support Vector Machine (SVM), K-Nearest Neighbour (K-NN) and Deep Learning models U-Net and Deep Lab version 3. Manual classification is used as the reference method.