Sketch and paint: A structural–temporal aware refinement network for fine-grained crop mapping on satellite image time series

Abstract

Crop type mapping based on Satellite Image Time Series (SITS) is a cornerstone of large-scale agricultural monitoring. However, existing methods often struggle in agricultural landscapes characterized by field fragmentation and crop phenological variability, where inter-class boundary ambiguity and intra-class phenological heterogeneity make it difficult to preserve sharp crop boundaries while maintaining spatial consistency. Inspired by the human cognitive process of “sketching outlines before filling in details”, we propose STAR, a sequential structural–temporal aware refinement framework for fine-grained crop mapping. STAR integrates three complementary components. The Structure Prior Injection (SPI) module explicitly extracts high-frequency geometric cues in shallow layers to anchor pixel-level boundary information and mitigate boundary degradation caused by deep feature diffusion. The Decoupled Phenology Encoder (DPE) module independently models pixel-wise growth trajectories, reducing spectral interference among phenologically similar crops and enhancing temporal discriminability. The Spatial Consistency Refinement (SCR) module further incorporates spatial context to enforce semantic coherence within structurally constrained regions without compromising boundary integrity. We evaluate STAR on four public crop mapping datasets covering study areas in France, Germany, and South Africa. Experimental results demonstrate that STAR consistently achieves state-of-the-art (SOTA) performance across multiple metrics. Qualitative analyses further confirm improved boundary delineation and field-level spatial coherence, highlighting the effectiveness of the proposed structural–temporal decoupling strategy. The source code is publicly available at https://github.com/cuiwei118/STAR .