Transformers for Remote Sensing: A Systematic Review and Analysis

Area of Science:

• Remote Sensing
• Computer Vision
• Artificial Intelligence

Background:

• Transformer models have seen increased research in remote sensing (RS) since 2021.
• A review of transformer applications in RS is currently lacking.

Purpose of the Study:

• To quantitatively analyze major research trends of transformers in RS over the past two years.
• To identify key application domains and performance characteristics of transformers in RS.

Main Methods:

• A quantitative analysis of transformer research in RS was conducted.
• Applications were categorized into eight domains: LULC classification, segmentation, fusion, change detection, object detection, object recognition, registration, and others.

Main Results:

• Transformers demonstrate higher accuracy in land use/land cover (LULC) classification and data fusion.
• Stable performance was observed in segmentation and object detection tasks.
• Transformers require more parameters than convolutional neural networks (CNNs) and need improved inference speed.
• Common application scenes include urban, farmland, and water bodies, primarily in natural sciences like agriculture and environmental protection.

Conclusions:

• Transformers offer significant potential in various RS applications, particularly in classification and fusion.
• Addressing parameter count and inference speed are crucial for advancing transformer performance in RS.
• Future research should focus on optimizing transformer architectures for RS data and expanding their application scope.