Palm and olive tree detection from UAV images with CNNs

Context and background Inventories of natural resources especially those of trees allow to make a systematic collection of information of existing trees in a given area. It therefore plays a crucial role in various fields such as forestry, forest fire control, climate change impact, and ecological studies. As such, they have become valuable data sources for decision-making regarding the management of natural resources. This is where Drone technology has been able to prove its efficiency in terms of cost and time. Along with DEEP Learning algorithms especially CNNs, they have been able to exceed traditional approaches, in time, cost, and precision. Goal and objectives This study explores the performance of two widely used algorithms for object detection. The comparison was conducted using two pre-trained CNN models commonly applied in UAV Remote Sensing: Ultralytics’ YOLOV5-Large and MobileNET SSD from the TensorFlow model zoo. Methodology The proposed approach for this study involves five main steps; creating a database, retraining both models on this database, evaluating their performance, and finally comparing their performance. Results With, respectively, learning and accuracy rates of 65% and 80% the objectives set by this study have been met. Both models were able to detect olive and palm trees from UAV images. This work could have been more efficient, if the training dataset was larger and included more palm trees annotations. Overall, YOLOV5-l showed better performance in palm and olive tree detection from UAV images. The training phase was faster, the detection of objects in the testing phase is also faster compared to SSD MOBILENET V2.