研究目的
To explore the feasibility of moving satellite imagery processing algorithms onboard satellites using COTS components, specifically for object classification such as ship detection, to enable real-time processing in space missions.
研究成果
The study demonstrates that advanced onboard processing for satellite imagery, such as ship detection using neural networks, is feasible with COTS hardware like the Nvidia Jetson TX2, fitting within SWAP constraints of SmallSats. This enables real-time applications in space, reducing reliance on ground processing and supporting future interplanetary missions. Future work should focus on code optimization and expanding to classify more object types.
研究不足
The code is written in Python using high-level libraries, which may not be optimized for embedded hardware; performance could be improved with CUDA or TensorRT. The algorithm is specific to ship detection on open ocean and may not generalize to land targets or other objects without modifications. Radiation tolerance of COTS parts is not fully addressed, posing risks for space environments.
1：Experimental Design and Method Selection:
The study uses a computer vision algorithm combining traditional techniques (edge detection and sliding window) with a convolutional neural network (CNN) for object detection and classification. The design rationale is to leverage COTS hardware for space applications.
2：Sample Selection and Data Sources:
Satellite imagery from Planet's Open California dataset (RGB or multispectral, max resolution 3 meters) and the Ships in Satellite Imagery dataset from Kaggle (700 ship images, 2100 not-ship images).
3：List of Experimental Equipment and Materials:
Nvidia Jetson TX2 embedded computer, Nvidia GTX 1080 Ti GPU for training, software libraries including Keras and TensorFlow.
4：Experimental Procedures and Operational Workflow:
Steps include applying Canny edge detection to satellite images, using a sliding window (80x80 pixels, stride 20) to generate image chips, filtering chips based on white pixel count (50-600 range), classifying chips with a CNN (VGGnet-inspired architecture), iterative training to reduce false positives, and performance testing on the Jetson TX2 in Max-Q and Max-N modes.
5：Data Analysis Methods:
Performance metrics measured include processing time, accuracy, false positive rate, and power consumption; statistical analysis of neural network accuracy and timing data.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
