Novel Sparse Adaptive Token Mamba Models for Remote Sensing Imagery Classification

Degree Program: MSc

Student’s Name: Zachary Dewis

Supervisor Name: Dr. Lincoln Linlin Xu

Thesis Title: Novel Sparse Adaptive Token Mamba Models for Remote Sensing Imagery Classification

Summary of Thesis:

Remote sensing image (RSI) classification, which transforms raw imagery into thematic maps, plays a critical role in environmental monitoring and resource exploration. However, achieving accurate and efficient RSI classification is a challenging task due to several key factors, including weak and subtle class signatures, difficulty in preserving fine spatial details, and the large volume of data. Existing deep learning approaches struggle to fully address these issues: Convolutional Neural Networks (CNNs) are limited in modeling long-range dependencies, while Transformer-based models suffer from high computational cost and potential overfitting. Although recent State Space Models, particularly Mamba architectures, offer a promising balance between efficiency and performance, they rely on fixed and rigid tokenization strategies that do not effectively capture the complex spatial, spectral, and temporal heterogeneity of RSI data.

To address these limitations, this thesis proposes a novel Sparse Adaptive Token Mamba framework, with five key principles: (1) reducing token quantity through sparsity to improve efficiency, (2) enhancing token quality by prioritizing spectrally pure representations, (3) dynamically learning token ordering based on importance, (4) enabling data-adaptive tokenization, and (5) designing disentangled architectures that separately model spatial, spectral, and temporal information. Together, these innovations improve class separability, particularly for weak spectral signatures, better preserve fine spatial structures, and significantly reduce the computation cost. 

The proposed framework is instantiated through three specialized models tailored to different data regimes: STSMamba for high temporal resolution data (e.g., MODIS time series), OBIA-Mamba for high spatial resolution multispectral imagery (e.g., Sentinel-2), and CSSMamba for high spectral resolution hyperspectral data. Extensive experiments across multiple benchmarks demonstrate consistent state-of-the-art performance. Overall, the framework provides a scalable and robust foundation for next-generation intelligent remote sensing analytics, with strong potential for multimodal data fusion and emerging sensor technologies.