Trends in Image Object Detection and Segmentation with Deep Neural Network on Natural and Earth Observations

Deep learning has a profound impact across multiple scientific domains and has increasingly positioned itself as a versatile tool for addressing new challenges in the field of natural (Earth) observation. However, researchers in Earth observation often face significant entry barriers, primarily due to the complexity and rapid evolution of the field, which is heavily driven by innovations in computer vision. To assist researchers in overcoming these challenges, this paper provides a comprehensive discussion of the development of deep learning, with a particular focus on image segmentation and object detection using convolutional neural networks. The paper begins with key developments when convolutional neural networks set new benchmarks in image recognition, and continues through to the innovations made recently. This paper traces the interconnections among the most influential convolutional architectures and major milestones originating from computer vision, facilitating a clear understanding of how these advances shape modern deep learning models. Additionally, we offer insights into the evolution of popular deep learning frameworks and present a summary of datasets commonly used in natural observation. By exploring well-performing architectures and evaluating their applications on Earth observation datasets, we assess how breakthroughs in computer vision influence future research in natural observation. This paper bridges the gap between theoretical advancements in computer vision and their practical implementation in natural observation, equipping researchers with the knowledge needed to effectively integrate deep learning into their work.