Editorial for the Special Issue “Advances in Object and Activity Detection in Remote Sensing Imagery”

Advances in data collection and accessibility, such as unmanned aerial vehicle (UAV) technology, the availability of satellite imagery, and the increasing performance of deep learning models, have had significant impacts on solving various remote sensing problems and proposing new applications ranging from vegetation and wildlife monitoring to crowd monitoring. This Special Issue contains seven high-quality papers [1,2,3,4,5,6,7] approaching problems relating to object detection, semantic segmentation, and multi-modal data alignment. In terms of the methods utilized, it is not surprising that six of the seven papers on this issue involve the application of deep learning. The papers also attest to the powerful aspect of the field where researchers can collaborate and validate their work on open-source models and datasets.

The first paper [1] addresses the problem of animal population estimation via thermal images, which often face the challenge of being low-resolution. The authors propose a modification to a popular object detection framework, naming it Distant-YOLO. The improved model, trained on a dataset containing low-resolution aerial images of rabbits, kangaroos, and pigs, was capable of detecting such animals, thus being potentially relevant for wildlife researchers and managers that previously relied on manual annotations.

The second paper [2] focuses on the detection of ships from synthetic-aperture radar images. The addressed problem relates to the fact that the accuracy of detection systems is often negatively affected by the complex background interference and the multi-scale features of ships. The authors propose the Quad-FPN architecture, a combination of four feature pyramid networks, which are all individually validated with extensive ablation studies. The work is potentially relevant for extremely important problems such as marine surveillance, traffic control, and fishery management. Likewise, the third paper [3] proposes an innovative architecture for the detection of objects through satellite optical imagery such as ships, but in a generalized manner. The authors address the problem that arbitrary objects in satellite imagery still pose a serious challenge for object detection models due to their diverse patterns in orientation, scale, and aspect ratio. The resulting model composed of such an active feature map realignment achieves higher performance, validated by the achievement of state-of-the-art results in two public datasets.

Semantic-segmentation-centered problems were also addressed with two papers presenting innovative enhancements. One semantic segmentation paper [4] aimed at urban vegetation cover estimation while addressing a problem often present in similar works given by an over-reliance on image color attributes. The improvement proposed is composed of a Multiview Semantic Vegetation Index (MSVI), which is implemented by a segmentation model (FCN and U-net) and with a proposed color mask adjustment. Given its multiview capability and ability to be applied to images such as those from Google panoramic cameras, the method has potential implications for real-time vegetation monitoring. In the second semantic segmentation paper [5], the subject approached was tidal flat waterbody estimation. Such a problem suffers from particular challenges where waterbodies differ little between their background while also contemplating blurry boundaries, which are difficult to detect accurately. As such a task represents one of the main ways to estimate waterbodies, it is increasingly relevant to aspects such as ecosystem protection and restoration, pollution control, and infrastructure construction.

The other two papers proposing novel solutions addressed the problems of crowd estimation and multi-camera space alignment. The innovative solution, presented by the paper [6] on crowd estimation by UAVs, aims at improving challenges in the form of the large requirement for the data required and onerous labeling by existing methods to obtain significant accuracy. Lastly, a paper [7] tackles the challenging task of aligning multiple cameras into a united coordinate system. In particular, the authors address the task of obtaining a cross-view between UAV deployed cameras and ground ones, creating an air-to-ground correspondence. The proposed solution is composed of methods that can create elaborate spatiotemporal feature maps and their cross-view space matching. This capability allows multiple cameras in a large-scale environment to be aligned into one coordination system with UAV auxiliary linkage. Therefore, such a development represents the relevant potential to enhance fields such as security surveillance, automatic control, and intelligent transportation.