Efficient Remote Sensing Image Classification with STConvNeXt
The ConvNeXt architecture takes its cue from the renowned ResNet model but stands distinct in its exclusive use of standard convolutional operations, deliberately steering clear of attention-based mechanisms. The framework unfolds over four hierarchical stages, each utilizing convolutional layers for spatial downsampling, while expanding channel dimensions. At its core lies the ConvNeXt Module, a multi-component structure consisting of depthwise separable convolutions, LayerNorm, channel expansion via convolutions, HardSwish activations, and residual skip connections. This amalgamation enhances representational capability without sacrificing computational efficiency.
ConvNeXt diverges from traditional ResNet architectures through three pivotal enhancements: larger convolutional kernels for broader receptive fields, substituting GELU for LayerNorm to bolster training stability, and employing depthwise separable convolutions to cut computational costs. The architecture smartly increases the channel expansion ratio to 4, augmenting feature extraction capacity.
Despite these advancements, ConvNeXt still holds two intrinsic limitations: significant parameter redundancy and limited feature abstraction capability. Addressing these concerns, we propose the STConvNeXt framework, which introduces two novel components: the SMConv module with spatial-channel decoupled operations and a tree-structured computation architecture for hierarchical feature recombination. These modifications curtail parameter complexity by 38% while bolstering representational power through multi-scale feature fusion.
The STConvNeXt framework comprises several key elements, as depicted in : the STConvNeXt block, fast spatial pyramid pooling (SPPF), global average pooling (GAP), and complete connectivity. This architecture is designed in a 3:3:9:3 configuration for feature extraction, followed by SPPF downsampling to decrease image size and computational complexity. Ultimately, features traverse through GAP, LayerNorm, and a fully connected layer for classification results.
In-depth within the STConvNeXt framework, the SMConv module effectively addresses redundancy by employing a branched structure inspired by architectures such as GhostNet and SSConv. The Representative branch uses convolutions to extract essential image features, capturing local features with small receptive fields and aggregating global information in deeper layers. Meanwhile, the Redundant branch supplements any lost information caused by channel separation. To further counteract redundancy, pointwise convolutions and depthwise separable convolutions balance the channel count in the Representative branch, whilst GAP merges the channels to recoup information lost to channel separation.
The mathematical formulation of SMConv is elegantly described in the equations provided, illustrating the sophisticated interworking of components such as group convolution, pointwise convolution, adaptive average pooling, and others within the module’s structure.
In the realm of CNNs, the strategic use of tree structures amplifies the model’s feature extraction capabilities and efficiency. The STConvNeXt block harnesses a tree structure for a multi-branch convolutional pathway, enabling the capture of features at varying scales, thereby enriching the model’s ability to decipher intricate patterns within remote sensing images.
Complementing the STConvNeXt framework is an innovative approach to addressing pooling’s tendency to lose vital visual information. Enhanced based on the YOLOV3 model, the improved SPPF method facilitates extracting features from diverse angles using fewer parameters than traditional convolutional downsampling, offering a balanced solution to this challenge.
The use of small versus large convolutional kernels within tree structures brings to light the delicate balance between accuracy and computational demands. Our study highlights how smaller convolutional kernels, when arranged in tree formations, can outperform their larger counterparts, achieving peak accuracy with slimmer parameters.
In pursuit of optimizing both accuracy and efficiency, we explored various configurations of convolutional kernels and their performance (as summarized in Table 1), ultimately identifying the tree structure as delivering the highest accuracy. Figure 6 further graphically illustrates the superior accuracy achieved by the tree structure, especially when utilizing the smallest kernel sizes.
Remote sensing image classification with STConvNeXt represents a pioneering stride forward, harmoniously blending sophisticated convolutional strategies with computational efficiency to redefine the landscape of remote sensing analytics.
