Spatial-Temporal-Spectral Mamba with Sparse Deformable Token Sequence for Enhanced MODIS Time Series Classification

TL;DR

This paper introduces a novel spatial-temporal-spectral Mamba model with deformable token sequences and a temporal grouped stem to improve MODIS time series classification, achieving higher accuracy and efficiency.

Contribution

It proposes a new deformable Mamba sequencing approach and a spatial-temporal-spectral architecture to better capture MODIS data features and reduce redundancy.

Findings

Higher classification accuracy compared to state-of-the-art methods

Reduced computational complexity in MODIS time series classification

Effective disentanglement of temporal-spectral features

Abstract

Although MODIS time series data are critical for supporting dynamic, large-scale land cover land use classification, it is a challenging task to capture the subtle class signature information due to key MODIS difficulties, e.g., high temporal dimensionality, mixed pixels, and spatial-temporal-spectral coupling effect. This paper presents a novel spatial-temporal-spectral Mamba (STSMamba) with deformable token sequence for enhanced MODIS time series classification, with the following key contributions. First, to disentangle temporal-spectral feature coupling, a temporal grouped stem (TGS) module is designed for initial feature learning. Second, to improve Mamba modeling efficiency and accuracy, a sparse, deformable Mamba sequencing (SDMS) approach is designed, which can reduce the potential information redundancy in Mamba sequence and improve the adaptability and learnability of the Mamba sequencing. Third, based on SDMS, to improve feature learning, a novel spatial-temporal-spectral Mamba architecture is designed, leading to three modules, i.e., a sparse deformable spatial Mamba module (SDSpaM), a sparse deformable spectral Mamba module (SDSpeM), and a sparse deformable temporal Mamba module (SDTM) to explicitly learn key information sources in MODIS. The proposed approach is tested on MODIS time series data in comparison with many state-of-the-art approaches, and the results demonstrate that the proposed approach can achieve higher classification accuracy with reduced computational complexity.