Introduction

Remote sensing helps us monitor the environment, respond to disasters, and improve security. Satellites and drones capture images that are analyzed to track forest changes, detect floods, and even spot aircraft or ships. A crucial aspect of this field is small-target detection, which means identifying objects that occupy only a few pixels in large images.

Detecting small targets like planes or boats in satellite images is tough because:

• They are tiny compared to the background.
• Noise in the images makes them harder to distinguish.
• Traditional detection methods often miss these objects.

AI-powered methods, such as deep learning models, have helped improve detection accuracy. However, most models rely on powerful GPUs to function, making them too energy-consuming for space applications, where low power consumption is essential.

That’s why researchers (Fang et al., 2025; Zhou et al., 2025) have explored FPGA architectures—a more efficient alternative to GPUs. FPGA-based models provide:

• Real-time image processing without needing to send data to the ground.
• Energy-efficient computation, making them suitable for satellites and drones.
• Faster response times compared to traditional remote sensing detection methods.

This blog explores how AI-powered FPGA models improve small-target detection in remote sensing, making real-time monitoring more efficient.

Understanding Small-Target Detection in Remote Sensing

Detecting small objects in satellite images is highly complex, primarily due to their minimal pixel footprint, interference from background noise, and changing environmental conditions (Li et al., 2025). Unlike large buildings or landscapes, small objects can be easily misclassified or missed entirely, leading to low detection accuracy.

Challenges in Detecting Small Targets

• Low image resolution: Objects appear blurry or distorted.
• Background interference: Clouds, terrain, and noise can hide small targets.
• High false detection rates: Standard models mistake irrelevant features for actual objects.

Traditional remote sensing relied on manual feature extraction, where experts studied images to identify targets (Wei et al., 2018). However, this method is slow and inconsistent, especially when dealing with complex environments.

AI-Based Small-Target Detection

Modern approaches use AI to automate the detection process, making it faster and more reliable. Methods like YOLO (You Only Look Once) work by:

• Scanning satellite images in real-time.
• Identifying patterns to recognize small objects accurately.
• Classifying detected objects with minimal human intervention (Wang et al., 2024).

Limitations of Existing Detection Models

Although deep learning has improved remote sensing, most models rely on high-performance GPUs, which require too much power for embedded satellite systems (Cook et al., 2018).

FPGA-based models offer a better solution, allowing small-target detection without relying on cloud processing, making real-time monitoring more practical for satellites and drones (Fang et al., 2025).

Methodology: The Proposed FPGA-Based Detection System

Detecting small objects in satellite images is a big challenge, especially when using traditional deep learning models that rely on power-hungry GPUs. These models work well in high-performance computing setups, but they struggle with real-time processing in space-based systems, where power and storage are limited.

To solve this, Fang et al. (2025) proposed an improved AI model called GF-YOLO, built specifically for small-target detection in satellite images. This model runs efficiently on FPGA (Field Programmable Gate Array) hardware, offering real-time processing with minimal power consumption.

GF-YOLO is designed to:

• Detect small objects more accurately while handling complex backgrounds.
• Reduce computational complexity, making it ideal for embedded systems.
• Run efficiently on FPGA hardware, offering lower power usage than traditional GPUs.

GF-YOLO: A Lightweight Model for Small-Target Detection

Traditional deep learning models struggle with small-target detection because they require a lot of computing power. GF-YOLO solves this issue by optimizing YOLOv5s, a widely used AI model, for low-power real-time detection (Zhou et al., 2025).

What makes GF-YOLO different?

Feature	Benefit
Lightweight architecture	Requires fewer resources, making it faster.
Optimized feature extraction	Improves accuracy when detecting small objects.
Works on FPGA hardware	Reduces power consumption while maintaining performance.

Instead of relying on GPUs, GF-YOLO runs on low-power FPGA processors, allowing satellites and drones to analyze images in real time (Chen et al., 2025).

GhostBottleneckV2 Module: Making GF-YOLO More Efficient

Deep learning models often require complex computations, which slows down processing. GF-YOLO improves efficiency by integrating the GhostBottleneckV2 module, which makes AI models lighter without losing accuracy (Li et al., 2025).

How does GhostBottleneckV2 help?

Component	Function	Impact
Ghost module	Generates extra features without complex operations.	Cuts down computation time.
Decoupled Fully Connected (DFC) attention	Focuses on small-object details.	Boosts detection accuracy.
Residual connection	Helps AI models learn better.	Reduces errors in predictions.

By replacing heavy convolution layers with GhostBottleneckV2, GF-YOLO runs faster and requires less processing power, making it perfect for satellite-based image analysis (Chen et al., 2025).

Remote Sensing: Hybrid Overlapping Acceleration Architecture for FPGA

One of the biggest innovations of GF-YOLO is its hybrid overlapping acceleration architecture, which optimizes AI processing on FPGA hardware. Traditional AI models rely on GPUs, which consume a lot of power. In contrast, FPGA-based acceleration offers efficiency and real-time processing (Fang et al., 2025).

How does hybrid overlapping acceleration improve performance?

Technique	Purpose	Impact on Processing Speed
Layer Fusion	Combines multiple layers to reduce complexity.	Speeds up detection time.
Double Buffering	Loads new data while processing the current batch.	Eliminates waiting delays.
Custom Processing Elements (PEs)	Dedicated AI hardware for key operations.	Makes detection smoother and faster.

With these improvements, GF-YOLO achieves:

• 67.8% average precision (mAP) when detecting small targets (Fang et al., 2025).
• Consumes only 2.8W of power, much less than traditional models.
• Processes images in real time, making it highly efficient for remote sensing tasks.

How the FPGA-Based AI Model Works

Detecting small objects in satellite images requires quick processing and low power usage since satellites can’t afford the heavy energy demands of traditional GPUs. To solve this, Fang et al. (2025) designed an FPGA-based AI model that processes images in real time while consuming much less power.

This section explains how the FPGA-based model works step by step, what optimizations make it efficient, and how it compares to GPU-powered solutions for remote sensing.

How the FPGA-Based AI Model Processes Images in Remote Sensing

Satellites and drones collect remote sensing images, but raw data must be cleaned, analyzed, and classified to detect important objects. Traditional methods rely on transmitting images to ground-based GPUs, which takes time and consumes a lot of power. The FPGA-based AI model processes images onboard, reducing delays and making real-time decisions easier.

Here’s the step-by-step process of how the FPGA model handles satellite images:

Step	Process	Impact on Detection
1. Image Acquisition	Captures thermal infrared or radar images from satellites or drones.	Provides raw data for detection.
2. Preprocessing	Cleans the image by removing noise and adjusting brightness.	Improves clarity and detection accuracy.
3. Feature Extraction	Identifies key patterns using the GhostBottleneckV2 module.	Helps the AI recognize small objects effectively.
4. Object Detection	Analyzes extracted features and classifies detected objects.	Determines whether the object is a valid target.
5. Post-Processing	Refines bounding boxes and removes false detections.	Ensures accurate identification of small targets.
6. Final Output	Sends detected object data for analysis or action.	Enables real-time monitoring without ground processing delays.

Since this model runs directly on the satellite hardware, it removes transmission delays, allowing faster and more reliable detection (Zhou et al., 2025).

Optimizations for Low-Power, High-Performance Detection

One of the biggest challenges of running AI models in space is power limitations. Traditional deep learning models need high-power GPUs, but satellites must operate efficiently while using minimal energy. Fang et al. (2025) solved this issue by implementing several optimizations in their FPGA-based GF-YOLO model to ensure high accuracy with minimal power usage.

Key optimizations include:

Optimization	Purpose	Impact on Performance
GhostBottleneckV2 Module	Reduces complexity while improving accuracy.	Allows faster processing with lower power.
Hybrid Overlapping Acceleration	Uses parallel computing to speed up object detection.	Improves processing speed without increasing power usage.
Layer Fusion	Merges neural network layers to reduce unnecessary computations.	Optimizes efficiency and minimizes delays.
Double Buffering	Loads new data while processing previous images.	Eliminates waiting time between frames.
FPGA-Optimized Hardware	Uses custom AI processing elements instead of generic computing units.	Lowers energy consumption and improves real-time processing.

These improvements allow satellites to process images locally instead of sending them to Earth, enabling instant decision-making for tasks like disaster monitoring and security surveillance (Li et al., 2025).

Comparing FPGA-Based Models to Traditional GPU-Based Solutions

Most AI-powered satellite detection models rely on GPUs, which offer strong computing power but are too energy-demanding for space applications. The FPGA-based model developed by Fang et al. (2025) solves this problem by delivering high detection accuracy while using much less power.

Feature	FPGA-Based GF-YOLO	GPU-Based Models
Energy Consumption	2.8W, ideal for embedded platforms (Fang et al., 2025).	Up to 90W, which is too much for satellites.
Processing Speed	Optimized for real-time small-target detection.	Powerful but limited by energy constraints.
Parallel Computing	Uses dedicated AI processing elements for efficient execution.	General-purpose processing, less optimized for specific tasks.
Deployment Feasibility	Works in low-power environments like satellites and drones.	Requires high-power infrastructure to run smoothly.
Inference Time	Faster due to hybrid acceleration techniques.	Can be slower when handling massive datasets.

Switching from GPUs to FPGA-based AI allows satellites and drones to process images instantly, without waiting for transmission delays (Zhou et al., 2025).

Results: Performance Comparison and Real-World Application

To test how well the FPGA-based GF-YOLO model detects small objects in satellite images, researchers ran several experiments on two datasets—TIFAD (Thermal Infrared Flying Aircraft Dataset) and HRSID (High-Resolution SAR Image Dataset). The results showed that GF-YOLO detects small targets better, runs efficiently, and uses much less power compared to older models like YOLOv5s and YOLOv4-tiny.

How GF-YOLO Performed on the TIFAD Dataset

The TIFAD dataset consists of thermal infrared images, which are useful for detecting aircraft in different weather conditions (Li et al., 2025). Since these objects appear as tiny heat spots, standard detection models often struggle to identify them accurately.

GF-YOLO vs. Older Models on TIFAD

Model	Detection Accuracy (mAP)	Number of Parameters (M)	Processing Load (GFLOPs)
YOLOv4-tiny	53.2%	5.91M	3.43 GFLOPs
YOLOv5l	62.3%	46.63M	57.31 GFLOPs
YOLOv5s	61.9%	7.07M	8.19 GFLOPs
YOLOv5s + GhostBottleneckV2	65.2%	5.81M	5.85 GFLOPs
GF-YOLO	67.8%	5.22M	4.87 GFLOPs

Key Takeaways from TIFAD Testing

• GF-YOLO detected aircraft more accurately, with 67.8% mAP, a significant boost over YOLOv5s.
• GhostBottleneckV2 helped the model detect tiny objects more efficiently while using fewer computing resources (Fang et al., 2025).
• GF-YOLO processed images faster than other models, making it ideal for satellites and drones that need real-time detection.

How GF-YOLO Performed on the HRSID Dataset

The HRSID dataset contains satellite radar images of ships, used to test how well AI models detect objects in varied lighting and environmental conditions (Wei et al., 2025).

GF-YOLO vs. Ship Detection Models on HRSID

Model	Detection Accuracy (mAP)	Number of Parameters (M)	Processing Load (GFLOPs)
SSD512	85.46%	24.4M	214.0 GFLOPs
YOLOv8s	90.1%	11.1M	14.2 GFLOPs
YOLOX-Tiny	87.72%	5.8M	11.9 GFLOPs
AMANet	91.4%	17.3M	24.7 GFLOPs
YOLOv8-FDF	78.8%	6.2M	12.3 GFLOPs
GF-YOLO	93.2%	5.22M	4.78 GFLOPs

Key Takeaways from HRSID Testing

• GF-YOLO had the highest accuracy (93.2%), beating YOLOv8s and YOLOX-Tiny.
• Uses fewer parameters than competitors, making it lightweight and energy-efficient.
• Works well in difficult environments, detecting ships even in complex scenes like ports.

Energy Efficiency: FPGA vs. CPU vs. GPU

One major advantage of GF-YOLO on FPGA is its low power consumption. Unlike GPUs, which require a lot of energy, FPGA models operate efficiently on satellites (Zhang et al., 2025).

Comparing Energy Usage Across Platforms

Platform	Processing Power (GOPs)	Power Usage (W)	Efficiency (GOPs/W)
Intel i9-13900 CPU	48.1 GOPs	56.2W	0.85 GOPs/W
RTX 4060 GPU	104.5 GOPs	90.5W	1.15 GOPs/W
FPGA-Based GF-YOLO	10.2 GOPs	2.8W	3.64 GOPs/W

Why FPGA Is Better for Remote Sensing

• Uses far less power than CPUs and GPUs, making it ideal for satellites with energy limits.
• Processes images efficiently, ensuring faster decision-making for real-time surveillance.
• Nearly 4x more efficient than traditional CPUs, proving its usefulness for low-power AI applications.

Conclusion: GF-YOLO’s Success in Remote Sensing AI

The FPGA-powered GF-YOLO model has proven itself as a fast, energy-efficient AI solution for remote sensing. Here’s what makes it stand out:

• Detects aircraft and ships more accurately than previous AI models.
• Uses less power, making it perfect for satellites, drones, and other embedded systems.
• Processes images quickly, allowing real-time object detection without delays.

By integrating lightweight optimizations, Ghost modules, and FPGA acceleration, GF-YOLO sets the new standard for small-object detection in remote sensing applications (Chen et al., 2025).

6. Future Innovations in Remote Sensing AI

AI is changing the way satellites analyze images, making remote sensing faster and more accurate. Researchers are now working on self-learning anomaly detection, privacy-protecting AI, and low-power processing to improve how satellites handle large amounts of data in space (Fang et al., 2025).

Smarter Anomaly Detection in Satellite Images

Most remote sensing models use fixed rules to spot unusual patterns, but they often miss unexpected targets. The new adaptive AI models can learn from real-time data, making them smarter over time (Li et al., 2025).

What’s new in anomaly detection?

• AI models that adjust in real-time to changing weather and lighting conditions.
• Better filtering of false alerts, preventing mistakes in target detection.
• Combining different image sources, like infrared and radar, for more accurate results (Zhou et al., 2025).

These updates will make satellites better at spotting aircraft, ships, and environmental changes with minimal human input.

How AI Protects Privacy in Remote Sensing

With AI processing millions of satellite images, security is more important than ever. Remote sensing systems must protect sensitive data while ensuring real-time analysis (Chen et al., 2025).

How AI is improving privacy:

• Federated learning, which lets satellites train AI models without storing raw image data.
• Encrypted image processing, where satellites analyze images without exposing them to hackers.
• Adding controlled noise to images, making it harder for unauthorized parties to access sensitive locations (Zhang et al., 2025).

These techniques make sure that satellites can process images without compromising security.

Using Edge Computing and Low-Power AI in Space

Satellites collect massive amounts of data every day, but sending everything to Earth for analysis wastes time and energy. Instead, AI models can process data onboard using low-power AI chips and edge computing (Fang et al., 2025).

Why this is important:

• Faster satellite image analysis, helping in disaster response and security tracking.
• Less reliance on Earth-based systems, making space technology more independent.
• Energy-efficient AI processing, allowing satellites to run AI models without draining their batteries.

FPGA-powered models like GF-YOLO have already shown success in real-time detection, paving the way for more power-saving AI applications in space (Zhou et al., 2025).

7. Conclusion: The Future of AI-Optimized Remote Sensing

AI-powered remote sensing is transforming how satellites monitor the Earth, making image processing faster, smarter, and more energy-efficient. By using low-power AI and FPGA acceleration, researchers are solving key challenges in space-based detection.

Why FPGA Acceleration Matters for Remote Sensing

FPGA-based AI models like GF-YOLO are changing the game in space-based object detection:

• Accurate detection of small aircraft, ships, and environmental changes.
• Consumes less power, making it ideal for long-term space missions.
• Processes images instantly, preventing delays in analysis (Chen et al., 2025).

FPGA acceleration ensures that AI models can run efficiently onboard satellites, reducing reliance on ground-based processing.

How AI Enhances Environmental Monitoring and Security

AI-powered remote sensing helps satellites:

• Respond quickly to disasters, like wildfires and floods.
• Track climate changes, such as glacier movements and forest loss.
• Improve security monitoring, detecting suspicious activity in global airspace (Fang et al., 2025).

With real-time AI detection, satellites can analyze images instantly, leading to better decision-making in environmental and security missions.

Next-Generation AI Models for Smarter Space Processing

The future of remote sensing will focus on AI models that use less energy while processing data faster. Researchers are working on:

• Smaller AI models, improving accuracy without using large computing resources.
• Advanced FPGA designs, allowing satellites to process images autonomously.
• Real-time AI automation, helping satellites make faster decisions without human input (Zhou et al., 2025).

These innovations will make space-based AI smarter, faster, and more energy-efficient, improving disaster response, global monitoring, and security tracking.

Click here to see more.

Reference

Fang, N., Li, L., Zhou, X., Zhang, W., & Chen, F. (2025). An FPGA-Based Hybrid Overlapping Acceleration Architecture for Small-Target Remote Sensing Detection. Remote Sensing, 17(494). https://doi.org/10.3390/rs17030494.

Creative Commons Attribution (CC BY) License

This article is licensed under the Creative Commons Attribution (CC BY) 4.0 License