A Multi-Scale Attention-Enhanced Architecture for Gravity Wave Localization in Satellite Imagery DOI: XXXXXXX.XXXXXXX Conference: Make sure to enter the correct conference title from your rights confirmation email; June 03–05, 2018; Woodstock, NY ISBN: 978-1-4503-XXXX-X/2018/06

Seraj Al Mahmud Mostafa, Jianwu Wang · 2025

Read A Multi-Scale Attention-Enhanced Architecture for Gravity Wave Localization in Satellite Imagery DOI: XXXXXXX.XXXXXXX Conference: Make sure to enter the correct conference title from your rights confirmation email; June 03–05, 2018; Woodstock, NY ISBN: 978-1-4503-XXXX-X/2018/06 by Seraj Al Mahmud Mostafa, Jianwu Wang (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2508-02704.txt.

Introduction

Seraj Al Mahmud Mostafa and Jianwu Wang build a detector for air waves in satellite images in a 2025 paper. Gravity waves in the upper air vary in size, shape, and span, which makes them hard to pin down. Clouds, city lights, and sensor noise crowd the same single-channel images.

Their model extends a known fast detector with two add-ons. One reads features at many scales through spread filters run side by side. The other weighs space and channel cues to stress true signal and mute noise. Scores beat rival models by wide margins on two standard tests. The gains matter for weather and climate watch from orbit.

Catalog record

CREATORSeraj Al Mahmud Mostafa, Jianwu Wang
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE48 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2508-02704

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

A Multi-Scale Attention-Enhanced Architecture for Gravity Wave Localization in Satellite Imagery DOI: XXXXXXX.XXXXXXX Conference: Make sure to enter the correct conference title from your rights confirmation email; June 03–05, 2018; Woodstock, NY ISBN: 978-1-4503-XXXX-X/2018/06 Seraj Al Mahmud Mostafa, Jianwu Wang Abstract . Satellite images present unique challenges due to their high object variability and lower spatial resolution, particularly for detecting atmospheric gravity waves which exhibit significant variability in scale, shape, and pattern extent, making accurate localization highly challenging. This variability is further compounded by dominant unwanted objects such as clouds and city lights, as well as instrumental noise, all within a single image channel, while conventional detection methods struggle to capture the diverse and often subtle features of gravity waves across varying conditions. To address these issues, we introduce YOLO-DCAT incorporating Multi Dilated Residual Convolution (MDRC) and Simplified Spatial and Channel Attention (SSCA), an enhanced version of YOLOv5 specifically designed to improve gravity wave localization by effectively handling their complex and variable characteristics. MDRC captures multi-scale features through parallel dilated convolutions with varying dilation rates, while SSCA focuses on the most relevant spatial regions and channel features to enhance detection accuracy and suppress interference from background noise. In our experiments, the improved model outperformed state-of-the-art alternatives, improving mean Average Precision (mAP) by over 14% and Intersection over Union (IoU) by approximately 17%, demonstrating significantly improved localization accuracy for gravity waves in challenging satellite imagery and contributing to more precise climate research and modeling. 1. Introduction Satellite data plays a vital role in Earth informatics by offering valuable insights into the planet’s environment and climate, providing critical information for interpreting a range of phenomena such as atmospheric conditions, sea surface changes, and environmental shifts. Additionally, satellite imagery is indispensable for studying gravity waves in the Earth’s atmosphere. Gravity waves are oscillations caused by buoyancy forces that are distinct from gravitational waves and play a crucial role in atmospheric dynamics and energy transfer ( Jovanovic 2018 ) . They significantly impact weather patterns, atmospheric composition, and climate systems, making their study essential for comprehensive climate understanding and modeling ( Alexander and Holton 1997 ) . Localizing gravity waves in the atmosphere is vital for climate research, as it helps quantify their effects on global atmospheric circulation and energy distribution ( Fritts and Alexander 2003 ) . However, despite the wealth of information provided by satellite datasets, they present significant complexities ( Guo et al. 2007 ) . Particularly, the satellite data poses substantial difficulties in gravity wave datasets, where various interferences, such as city lights, clouds, and instrumental noise obscure subtle atmospheric phenomena ( Mostafa et al. 2025 ; González et al. 2022 ) . These datasets also encounter unavoidable challenges including significant variability in the scale, shape, and extent of the main object patterns. Gravity waves can be mixed with or hindered by interference from occlusion and overlap, complicating the detection process. Moreover, the fact that gravity wave datasets are captured and stored in a single band makes it challenging to distinguish and remove unwanted objects. The relative infrequency and low significance of gravity waves in these datasets further complicate their detection ( Curtis Seaman 2013 ) . To tackle these challenges and improve gravity wave localization, we propose YOLO-DCAT (YOLO with D ilated C onvolution and A ttention-aided T echnique), which provides enhancements to the existing state-of-the-art YOLOv5 object detection model ( Jocher et al. 2020 ) . Our approach focuses on two key contributions as follows: 1) Multi Dilated Residual Convolution (MDRC): This modification to the network’s backbone increases the receptive field without losing spatial resolution. The multi-dilation approach applies different dilation rates to convolutional layers, enabling the capture of features at various scales simultaneously. The residual approach ( He et al. 2016 ) with skip connections allows the network to learn residual functions, facilitating gradient flow and improving training efficiency. Feature fusion combines information from different dilation rates, enhancing the model’s ability to detect variable-scale gravity wave patterns effectively. 2) Simplified Spatial and Channel Attention (SSCA): Implemented after each Dilated Residual Block in the backbone, this mechanism enhances important features while suppressing noise interference. It focuses on relevant spatial areas and channels simultaneously, improving the model’s ability to distinguish gravity waves from background interference and unwanted objects. Labeled Data Y.+MDRC+SSCA CBAM ( Woo et al. 2018 ) ViT ( Dosovitskiy et al. 2020 ) Transformer ( Vaswani 2017 ) YOLO ( Jocher et al. 2020 ) Figure 1. Comparing the localization of Gravity Waves using various baseline and state-of-the-art models, including the proposed YOLO+MDRC+SSCA (denoted as Y.+MDRC+SSCA in the second column) model. The leftmost column represents the labeled data, where the object of interest (Gravity Waves) is highlighted in pink shades, while non-objects of interest (such as city lights, and clouds) are highlighted in blue shades. The top row displays data with more visible gravity wave occurrences, whereas the bottom row presents more challenging scenarios where gravity waves are partially obscured, blurred, or embedded within clouds. (Please refer to supplementary material for more results). These enhancements enable our model to better handle the challenges of noisy satellite data and improve the accuracy of gravity wave localization, contributing to more precise climate research and modeling. In Figure 1 , we present results demonstrating the effectiveness of our proposed model compared to other state-of-the-art models. The detection results show that our approach delivers more promising performance than both baseline and state-of-the-art models, with improved localization accuracy and higher confidence scores. A detailed discussion of the experimental results is provided in Section 5 . The structure of this paper is organized as follows. Section 2 provides an overview of the Gravity Wave dataset and the YOLO model architecture. Section 3 reviews related work in dilated convolutions and attention mechanisms. In Section 4 , we present the proposed methodology with detailed explanations of MDRC and SSCA components, followed by the experimental setup and comprehensive results in Section 5 . Section 6 offers a thorough discussion of our findings and their implications, and finally, Section 7 concludes this study with future research directions. 2. Background 2.1. Gravity Wave Dataset This study utilizes data from the night band of the Visible Infrared Imaging Radiometer Suite (VIIRS) Day/Night Band (DNB) on the Suomi NPP satellite ( Curtis Seaman 2013 ) . The VIIRS DNB captures broadband upwelling radiance in the visible spectrum, with a swath of approximately 3000 km and spatial resolution of about 1 km at nadir. Each 6-minute granule, typically measuring around 4000 \times 3000 pixels, is stored in Hierarchical Data Format version-5 (HDF5) and records radiance measurements in the 0.5–0.9 \mu m wavelength range. These images capture extremely low radiance levels, on the order of 10^{-9} W/ cm^{2} sr -1 ( Murphy et al. 2006 ; NOAA Gravity Wave Data [n. d.] ) , and face significant challenges due to various disturbances from city lights, clouds, and inherent instrumental noise, which are further compounded by limited availability of ground truth annotations. To highlight the airglow associated with gravity wave events, nighttime images taken under new moon conditions are utilized, following established protocols described in ( Miller et al. 2015 ; Hu et al. 2019 ) . The preprocessing pipeline involves several critical

Read the complete text