Efficient remote sensing image classification using the novel STConvNeXt convolutional network.

Area of Science:

• Computer Vision
• Remote Sensing
• Machine Learning

Background:

• Remote sensing image classification faces challenges from complex spatial structures, high inter-class similarity, and intra-class variability.
• Existing methods often struggle to balance computational efficiency with effective feature extraction.

Purpose of the Study:

• To propose an innovative lightweight convolutional network, STConvNeXt, for efficient and accurate remote sensing scene classification.
• To enhance feature extraction and classification performance while minimizing computational resources.

Main Methods:

• Developed STConvNeXt featuring a split-based mobile convolution module with a hierarchical tree structure.
• Incorporated parameterized depthwise separable convolutions and a fast pyramid pooling module for efficient feature fusion and context awareness.
• Introduced a dynamic threshold loss function with a learnable inter-class margin to improve discrimination of challenging samples.

Main Results:

• STConvNeXt reduced parameter count by 56.49% and FLOPs by 49.89% compared to the ConvNeXt baseline.
• Achieved improved classification accuracy of 1.2-2.7% over the baseline.
• Demonstrated superior performance against state-of-the-art remote sensing scene classification models.

Conclusions:

• STConvNeXt offers a computationally efficient yet highly effective solution for remote sensing image classification.
• The proposed architectural modules and training strategy significantly enhance classification performance.
• The model maintains excellent accuracy despite substantial reductions in parameters and computational load.