Multi-Scale Patch Transformer Network for Satellite Image Time Series Classification
TL;DR: a multi-scale patch transformer model and for satellite image time series classification is proposed.
Abstract: With the increasing availability of high-quality earth observation data, satellite image time series (SITS) classification has become a hot topic. In this paper, a multi-scale patch Transformer model (PatchSITS) for SITS classification is proposed. First, SITS samples are segmented into patch sequences of varying patch lengths using k-means clustering. Subsequently, the enhanced Transformer is proposed to capture temporal features at various scales. To capture inter-band relationships and enhance critical band information, the gated channel attention mechanism is applied to obtain dynamic weights between bands. Furthermore, a multi-scale weighted fusion strategy is proposed to integrate these multi-scale features. And broad learning system (BLS) is utilized for SITS classification. Numerical simulations demonstrate that the PatchSITS model exhibits excellent classification performance on the BreizhCrops dataset.
Submission Number: 133
Loading