Deep Learning Algorithms for Remote Sensing
Date
2024
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
university of bordj bou arreridj
Abstract
Deep learning has revolutionized the analysis of data collected from unmanned aerial
vehicle (UAV) imagery, allowing for more profound insights, precise analysis, and enhanced
data extraction. This advancement has significantly contributed to the refinement of semantic
segmentation techniques. Particularly convolutional neural networks (CNNs) have emerged as
powerful tools in this domain, outperforming traditional methodologies. Nonetheless,
challenges persist, including feature extraction, class imbalance issues, overfitting, and
vanishing gradients that hinder deep neural network training, consequently impacting
segmentation performance. To address these challenges, we propose a novel approach by
integrating the U-Net architecture with ResNet34 backbone leveraging its strong feature
extraction capabilities. These features are further improved by using trained weights from the
ImageNet dataset. We train and evaluate the proposed model on several UAV datasets, including
Aerial Semantic Segmentation, LandCover.ai, UAVid, and AeroScapes. We achieve remarkable
performance, higher accuracy, precision, recall, F1-score, and miou compared to other methods.
Description
Unmanned Aerial Vehicles (UAVs) have revolutionized remote sensing, offering
continuous access to high-resolution data from previously inaccessible areas. This data is
particularly valuable for semantic segmentation, a deep learning technique crucial for tasks like
land cover mapping and object detection in environmental monitoring and urban planning.
However, effectively analyzing this data can be hampered by the vanishing gradient problem,
where information gets progressively lost during training in deep neural networks.
This research addresses this challenge by leveraging deep learning and specifically
tackling the vanishing gradient problem. We integrate a U-Net architecture with a ResNet34
backbone. The ResNet architecture is a deep learning model specifically designed to mitigate
the vanishing gradient problem. This allows the model to learn complex features within UAV
imagery for semantic segmentation tasks with greater depth and effectiveness. We trained and
evaluated the model on diverse UAV datasets (LandCover.ai, Aerial Semantic Segmentation,
UAVid, AeroScapes) leveraging GPUs, the Adam optimizer, and the Dice loss function. We
meticulously optimized hyperparameters and implemented specific training strategies to
maximize performance.
Our approach achieves remarkable results, with Intersection over Union (IoU) scores
exceeding 90% on LandCover.ai and AeroScapes datasets (92.0% and 91.0%, respectively).
Compared to existing methods (DeepLabv3, FCN-8, RCCT-ASPPNet, SS-Inhi-VGG16-FFT),
our model demonstrates superior accuracy across all metrics (precision, recall, F1-score,
mIoU). These findings solidify the power of deep learning for accurate and efficient large-scale
geospatial analysis in remote sensing.
Looking ahead, this research paves the way for further exploration of deep learning and
UAV data integration for semantic segmentation tasks. Future work can refine the proposed
approach, investigate new applications, and validate its robustness in real-world scenarios
Keywords
: Remote Sensing, UAV, Semantic Segmentation, Deep Learning, U-Net, ResNet34