A CNN-based algorithm for selecting tree-of-interest images acquired by UAV

• Published in: 2021 IEEE International Conference on Machine Learning and Applied Network Technologies (ICMLANT) pp. 1 - 6
• Main Authors: Salazar-Reque, Itamar, Arteaga, Daniel, Huaman, Kevin Guerra, Huaman Bustamante, S.
• Format: Conference Proceeding
• Language: English
• Published: IEEE 16.12.2021
• Subjects: CNN; Forestry; Geology; Image segmentation; image selection; Machine learning; Machine learning algorithms; Metadata; segmentation; Tree; UAV; Vegetation
• DOI: 10.1109/ICMLANT53170.2021.9690556

In recent years, the rapid development of unmanned aerial vehicles (UAV) and high-resolution cameras offered an important source of fine-grained imagery. This rich information has the potential to be used in many agricultural and forestry applications. Many of these applications require the monitoring of individuals trees along time. Thus, a crucial step is to select the best images of a specific tree from a big number of images acquired by the UAV. If made by hand, this is a very time-consuming activity and using geolocation metadata is not enough to tackle this problem. In this work, we present an algorithm to automate this process. The algorithm uses image geolocation data as a first filter for selections and tree segmentations to refine and select the best images. We used a convolutional neural network (CNN) to generate tree segmentations which achieved an accuracy of 0.98 when compared with manually segmented images. To test the image selection algorithm, we collected a total of 4807 RGB images in six different flights over an agricultural field with 144 avocado trees. We compare the selection algorithm outcomes with human selections per tree. The algorithm achieved an average true positive rate (TPR) of 0.88 for the selection of the three best images.