Deep Learning Enables On-board Hyperspectral Detection: A Key Path to Improve Target Detection Accuracy

On-board hyperspectral images are due to theirHigh spectral resolutionandLarge-scale spatial coverageHyperspectral targets have significant applications in tasks such as ground feature identification, urban monitoring, agricultural surveys, and environmental assessment. However, this type of data typically features numerous spectral bands, large target scale variations, complex backgrounds, and scarce labeled samples, making target detection tasks highly challenging. Traditional spectral matching and statistical feature methods struggle to simultaneously process spectral and spatial information, resulting in limited effectiveness for detecting small targets and complex backgrounds. In recent years, the development of deep learning technology has provided a new solution for on-board hyperspectral target detection. By automatically extracting spectral-spatial features through end-to-end training, it significantly improves the accuracy and robustness of detection.

Figure 1Examples of hyperspectral images from Xiguang-1 01 satellite

In terms of methodology,Spectral-spatial convolutional networks(like3D-CNNOne of the most common approaches is to use a hybrid CNN (or a combination of both). These methods treat hyperspectral data as...3D volume dataBy simultaneously performing convolutions in both spatial and spectral dimensions, it is possible to capture the continuity and spatial structure information between spectral bands. For example, 3D-CNN methods for crop identification tasks can effectively detect small-scale targets and improve overall detection accuracy through multi-layer convolutional feature fusion, but their computational cost is high and they are sensitive to the number of training samples. To highlight target features and suppress interference from complex backgrounds,Spectral-spatial attention mechanismThis approach is incorporated into the network structure. Through attention weighting in both spectral and spatial dimensions, it enhances the response in key bands and regions, proving particularly effective for small target detection, while also improving the model's robustness in complex environments.

Figure 2 shows the CNSST network, which uses a combination of parallel convolutional branches and Transformer branches to achieve hyperspectral target recognition.

Multi-scale feature fusionThis is another key technology for addressing target scale variations in satellite hyperspectral images. By integrating feature information from different convolutional layers or pyramid structures, the model can handle both large-scale and small-scale targets, achieving good detection results in scenarios such as urban buildings, roads, or farmland. For the problem of limited labeled data, transfer learning and few-shot learning are widely used. By pre-training the model on large-scale remote sensing data or other hyperspectral datasets and then fine-tuning it for a specific task, the problem of sample scarcity can be alleviated, training convergence speed can be accelerated, and high detection accuracy can be maintained.

Figure 3 shows MsFNet, which employs a multi-branch parallel architecture and combines dilated convolution, attention mechanism, and cross-scale feature interaction to achieve multi-scale spectral-spatial dynamic fusion in hyperspectral detection.

Despite the significant progress made by the above methods in on-board hyperspectral target detection, some challenges remain, including insufficient real-time processing capabilities for large-scale data, adaptability issues to different on-board platforms, and the difficulty of detecting small samples and rare targets.

Future development directions include lightweight model design for real-time detection, multimodal data fusion to enhance detection accuracy, self-supervised or semi-supervised learning to reduce reliance on labeled data, and Transformer-based network structures to capture global spectral-spatial dependencies. Overall, deep learning methods have become a core technology for on-board hyperspectral target detection. Through continuous optimization of network structures and learning strategies, it is expected to further improve detection accuracy and applicability, providing strong support for remote sensing monitoring and surface target identification.