Deep Learning-Based Enumeration of Pirogues Using Google Earth Images: A Case Study of Artisanal Fishing Landing Sites in Senegal

Satellite image analysis has gained paramount importance across diverse applications, ranging from land cover mapping to urban planning, disaster management, and environmental monitoring. Within this spectrum, the detection and enumeration of specific objects, such as pirogues (traditional fishing boats), hold significant value. However, object detection from satellite images is faced with challenges, including varying resolutions, occlusion, scale variations, and complex backgrounds. This research addresses a critical challenge in obtaining consistent and reliable data on the numbers of pirogues at fishing landing sites, a task often hindered by resource-intensive and spatially limited traditional data collection methods. The study aims to bridge this gap by introducing an artificial intelligence (AI) workflow that leverages Google Earth images to enumerate pirogues at artisanal fishing landing sites in Senegal. The workflow encompasses crucial stages, including data collection, annotation, model development, and robust analysis. 147 High-resolution Google Earth images serve as the primary dataset, undergoing annotation, preprocessing, and data augmentation to enhance diversity. Additionally, transfer learning, a technique utilizing pre-trained models on large datasets, is employed to enhance the accuracy and effectiveness of object detection. The performance of two well-known object detection algorithms, YOLOv5 and YOLOv8, is evaluated using the annotated data. Results indicate the superiority of YOLOv8, achieving a mean average precision at IOU 0.5 (mAP50%) of 70.5% and speed of 0.2 ms. Furthermore, integrating YOLOv8 with multiple object tracking facilitates the automation of pirogue counting at the Kayar fishing landing site.