Crop Anomaly Identification with Color Filters and Convolutional Neural Networks

Monitoring crops is a time consuming yet important task to ensure production quality. In this paper we present a comparison of convolutional neural network-based methods, evaluating model complexity and performance based on multiple metrics for the binary classification task of segmenting trees from the environment. An Unmanned Aerial Vehicle (UAV) is used to obtain RGB video of orange crops in different altitudes. Keyframes are extracted based on drone trajectory and speed for training and evaluation of the models. The effect on performance of multiple data augmentation techniques is also evaluated. The preferred model is then applied to a reconstruction of a region from multiple images and a color filter is applied for anomaly detection. Experimental and visual results show that these methods are able to segment the environment efficiently without any feature engineering, being a viable pre-processing method for reducing noise in disease identification applications.