Abstract:

Convolutional Neural Networks (CNNs) have been used in remote sensing applications, such as marine surveillance, traffic management, or road networks detection. However, since CNNs have extremely high computational. All this work has been conducted within the EU-funded Video Imaging Demonstrator for Earth Observation project. As it will be presented in this article, the work includes the introduction of a methodology based on the project constraints, the evaluation of different state-of-the-art CNN architectures by means of a new efficiency measurement also proposed in this work, the introduction of a new efficient CNN architecture, and finally, its optimized hardware implementation by means of high-level synthesis tools. The results obtained following the proposed methodology demonstrate that the uncovered architecture is able to detect targets of interest in RGB images with a much higher efficiency than state-of-the-art solutions, while requiring a much smaller amount of computing and memory resources.

Index Terms—Convolutional neural networks (CNNs), deep learning, machine learning, remote satellite images,target detection.

Introduction:

NOWADAYS, there is a rapid surge of interest in deep learning for remote sensing. These algorithms have rapidly become a hot topic in the machine learning field, and are currently
Manuscript received November 30, 2021; revised March 11, 2022; accepted April 12, 2022. Date of publication April 21, 2022; date of current version May 19, 2022. This work was supported in part by the European Union’s Horizon 2020 research and innovation program through Video Imaging Demonstrator for Earth Observation (VIDEO) project under Grant 870485 and in part by the SpanishGovernment and European Union (FEDER funds) through HypErsPEctRal
Imaging for Artificial intelligence applications (TALENT-HExPERIA) project under Grant PID2020-116417RB-C42. (Corresponding author: Sebastián López.)
The authors are with the Institute for Applied Microelectronic (IUMA),University of Las Palmas de Gran Canaria, 35001 Gran Canaria, Spain (email:rneris@iuma.ulpgc.es; armolina@iuma.ulpgc.es; rguerra@iuma.ulpgc.es seblopez@iuma.ulpgc.es; roberto@iuma.ulpgc.es).Digital Object Identifier 10.1109/JSTARS.2022.3169330 part of the remote sensing Big Data analysis paradigm. In particular, there has been a lot of progress made in the deep learning models, particularly convolutional neural networks (CNNs),which have significantly improved the performance of remote sensing image processing tasks. Nevertheless, the large number of model parameters and the high computational cost of CNN algorithms. Issues like resource utilization, minimal precision arithmetic using fixed-point representations, descriptions or real-time processing requirements are to be faced consumption, restricted computational capabilities, and limited storage availability.

Hence, in order to select the right architecture, a wide evaluation of existing CNN models has to be carried out. The work presented in this article has been conducted within the Video Imaging Demonstrator for Earth Observation(VIDEO) project,in which the aforementioned restrictions are of the utmost relevance. As it is shown in Section II, this project presents specific constraints that have to be taken into account, such as the large size of the images captured by the sensor, which impacts the amount of memory needed, or the encoding method used by the sensor, which will determinate the availability of the data.Within this context, them in contributions of this work are summarized as follows. First, the scanning algorithm used to process the entire image as independent blocks as soon as they are sensed has been evaluated according to different parameters, such as the sliding window size and stride. Second, different CNN architectures have been analyzed measuring both the detection performance and the computational cost, selecting the proposed-by-the-authors MobileNetv1Lite architecture as the most suitable option for this work according to a new figure of merit also uncovered in this article.