.png){kind=icon}
.png){kind=icon}
Return
Fine-grained crop classification from satellite image time series by boosting rare-class representations and enforcing global semantic consistency
.png){kind=icon}
OA
AI
A
J
Z
J
W
DOI：10.3389/frsen.2026.1822070.png)
.png){kind=icon}
.png){kind=icon}
Share.png){kind=icon}
Save
Abstract
En 中文
Accurate and timely fine-grained crop type classification from satellite image time series is crucial for large-scale agricultural monitoring and decision support in food-security management. However; fine-grained classification remains challenging due to extreme class imbalance and high inter-crop spectral similarity; especially when rare crops occupy only small and fragmented parcels. We propose a rare-class-aware framework with global semantic consistency regularization for fine-grained crop classification from Sentinel-2 multispectral time series. Built on a spatiotemporal encoder–decoder backbone; the framework combines rare-class-aware patch sampling with spatiotemporal perturbations to strengthen minority-class representations; and a global semantic consistency regularization based on patch-level class proportion estimation to align patch composition with pixel-wise predictions. Experiments on the H2Crop benchmark (France; 2022–2023); which contains over 1 million annotated parcels and 101 fine-grained crop types; validate the proposed strategy. Our method achieves a mean F1-score of 36.72% and an IoU of 27.82%; consistently outperforming state-of-the-art approaches; with improvements of 4.38 percentage points in F1-score and 4.00 percentage points in IoU over the strongest competitor. These results demonstrate strong potential for reliable rare-crop recognition and fine-grained agricultural monitoring in large; highly imbalanced landscapes.
Keywords:
class imbalance
sentinel-2
crop classification
multi-temporal
multitemporal remote sensing
rare crop types
AI Summary
Key information extracted from the uploaded paper, including a brief overview, abstract, background, key highlights, visual analysis, and future outlook.
Journal
F
IF:
3.7
Papers:
560
Citations:
993
Organization
Cited Papers