Multiscale Spectral–Temporal Fusion Network for Crop Mapping Using Sentinel-2 Time-Series Image

Abstract

Satellite image time series (SITS) from Sentinel-2 provide valuable temporal information for crop-type mapping. However, most existing approaches model crop phenology at a single temporal scale and may miss complementary short-term, seasonal, and annual dynamics. To address this limitation, this letter proposes a multiscale spectral–temporal fusion network (MSTFN) for pixelwise crop classification from Sentinel-2 time series. MSTFN uses Conv1D for local variations, BiLSTM for seasonal dynamics, and Transformer–BiLSTM for annual dependencies. An interscale self-attention module adaptively fuses these representations to support crop discrimination under overlapping temporal–spectral signatures. On the TimeSen2Crop benchmark, MSTFN achieves 87.85% overall accuracy (OA) and 86.81% weighted $F1$ -score (WF1), showing competitive-to-improved performance over representative CNN, recurrent, and attention-based baselines.