Coffee Crop Detection from UAS Orthomaps with Convolutional Neural Networks

This paper proposes a methodology for the automatic detection of coffee plants from close-range aerial images and deep learning techniques based on the YOLO framework. The data collection methodology, image preprocessing, and neural network training are discussed. Image datasets were collected at different locations in the central region of Costa Rica between 2018 and 2021. Models developed using this methodological approach had average mean accuracies up to 93% (mAP@0.35) on independent test sets; the detection performance achieved was 92% by using yolov4ext-ghost as a network architecture.