Deep Learning Applications on Very High-Resolution Aerial Imagery

Abstract

Very high-resolution (VHR) aerial imagery can capture finer details of information on the earth's surfaces. The availability of VHR aerial sensors has brought new opportunities and challenges in the development of advanced image-processing algorithms. Comparatively, aerial remote sensing is new in earth observation applications and has drawn a lot of research interest, as VHR aerial datasets have been effectively utilized in natural resource mapping, disaster management, surveillance, etc. The major issues with the VHR aerial images is that the traditional most advanced machine learning (ML) algorithms like support vector machine (SVM), artificial neural network (ANN), and random forest (RF) have failed to produce satisfactory results because of the large volume of data with higher feature dimension and the limited size of training samples. Deep learning (DL) is a subfield of machine learning that relies on automatic feature learning and classification. A recent advancement is due to an increase in computational power, the use of graphical processing units, and the development of efficient algorithms and activation functions. DL-based methods have produced a promising performance and revolutionized the image-processing activity in recent times. Significant efforts have been reported in the development of DL-based algorithms. Convolution neural network (CNN)-based methods have been found very useful in an image-based analysis such as object detection and semantic segmentation. For sequence or temporal analysis, recurrent CNN (RCNN)-based methods have proven to be more effective than the CNN approaches. Most of the DL-based approaches have been popularly utilized in event prediction and forecasting, data synthesis, object tracking, 3D semantic segmentation, image super-resolution, etc. This chapter focuses on the potential of DL techniques in realizing various image-processing applications. It summarizes the recent trends of DL techniques in the context of earth observation applications highlighting the issues, challenges, and limitations. It also demonstrates a case study on the automatic extraction of roads from VHR unmanned aerial vehicle (UAV) imagery using the DL technique. Comparative discussion on the currently available DL algorithms in terms of object detection and pattern recognition is provided in detail. The chapter concludes with the recommendations and way forward toward the utilization of DL techniques for various applications.