Recently, the demand for satellite image analysis increased with the recent advancement in various research areas (e.g., improvements in remote sensing image resolution, object recognition ideas, and deep learning techniques). Articles report new trends in object detection using remote sensing images, such as multi-temporal scene classifications, semantic segmentation, multiresolution, and large-scale optical image analysis. However, this study observed a black box of synchronization among the advancements in various technologies (i.e., new ideas in satellite image analysis) and advancements in deep learning techniques. Without investigating this black box, the optimization of object detection will be out of track. So, to keep the deep learning innovation on track, the primary goal of the current work is to explore the aforementioned black box to advance object detection in remote sensing images. In contrast, focusing on a specific technique, this study explores the black box while reviewing 150 articles from 2019 to 2023. First, the study highlights research methodologies and novel features of literature to achieve object detection in remote sensing images. Second, it evaluates effective deep learning models and assesses various featured studies to draw a clear picture of limitations and recent object detection trends to provide in-depth recommendations and future directions.
Top1. Introduction
The rapid increase in global security issues, urbanization, and natural disasters have created numerous societal challenges. On the one hand, the community needs an autonomous land surface monitoring system for various problems, such as human rights violations, border conflict management, and terrorist attacks in different countries (Al-Bilbisi, 2019; Charrua et al., 2021; Vogels et al., 2019; Levin et al., 2020). On the other hand, urbanization and environmental changes lead to drought, ultimately increasing plant mortality, reducing agriculture production, and causing various diseases in crops (Wang et al., 2021; Mzid et al., 2021; Alhichri et al., 2021; Wen et al., 2021; Olson and Anderson, 2021; Yi et al., 2021; Osco et al., 2021). Articles produce new trends and deep strategies in remote sensing data analysis to address these issues. (Vivekananda et al., 2021; Ru et al., 2020; Du et al., 2019; Xu et al., 2019; Wang et al., 2019). However, extensive literature studies depict that some models are good in prediction accuracy, while few are significant in rapid image analysis with low accuracy. These studies show that deep learning models are insignificant during satellite image analysis, even though they produce good results while evaluating classical photography or aerial imagery. Satellite images have very high resolutions, while classical photography has a relatively small amount of pixels. Also, classical images have red, green, and blue colors, while satellite images can have multiple red, green, and blue channels. The object is very small in the satellite image, consisting of a few pixels (Signoroni et al., 2019; Gu et al., 2019; Zhang et al., 2019; Silver et al., 2019; Pan et al., 2020). Satellite image has specific data with one shot, such as dealing with rain, cloudy, sunny, and different satellite positions. Also, we have to assess the Geo-referencing of an image. It means that we must know the object’s location in the picture. While evaluating remote sensing satellite images and various deep object detection strategies, we observed a black box of synchronization among the advancements in different technologies. The three main points of this black box are given below.
- •
Technologies for climate change monitoring.
- •
New ideas in satellite image analysis.
- •
Advancements in deep learning techniques.
The aforementioned extensive literature analysis shows that without investigating the black box, the optimization of object detection will be out of track. So, the current study proposes two basic questions to explore the black box and to keep the deep learning innovation on track. These questions are given below.
- •
Q1: What are the trended in-depth strategies to achieve object detection? How the particular object in conventional photography and remote sensing image is further explored?
- •
Q2: What are the limitations of trended deep learning models while processing remote sensing images?