Yang et al. (2026) From Sparse to Refined Samples: Iterative Enhancement-Based PDLCM for Multi-Annual 10 m Rice Mapping in the Middle-Lower Yangtze

Identification

• Journal: Remote Sensing
• Year: 2026
• Date: 2026-01-08
• Authors: Lingbo Yang, Jiancong Dong, Cong Xu, Jingfeng Huang, Yan Wang, H. Lü, Z. Chen, Wang Le
• DOI: 10.3390/rs18020209

Research Groups

• School of Artificial Intelligence, Hangzhou Dianzi University, Hangzhou 310018, China
• New Zealand School of Forestry, University of Canterbury, Private Bag 4800, Christchurch 8140, New Zealand
• Institute of Applied Remote Sensing and Information Technology, Zhejiang University, Hangzhou 310058, China
• Digitization and Informatics Division, Food and Agriculture Organization of the United Nations (FAO), 00153 Rome, Italy
• Institute of Agricultural Resources and Regional Planning, Chinese Academy of Agricultural Sciences, Beijing 100081, China

Short Summary

This study developed a Progressive Deep Learning Crop Mapping (PDLCM) framework to address sample scarcity and environmental heterogeneity for national-scale, high-resolution rice mapping. The framework successfully produced 10 m multi-annual rice maps for over 1,000,000 square kilometers in the middle and lower Yangtze River Basin from 2022 to 2024, achieving an average overall accuracy of 96.8% and an F1 score of 0.88.

Objective

• To develop a scalable and data-efficient deep learning framework for large-scale, high-resolution rice mapping in the Yangtze River Basin, aiming to reduce dependence on extensive labeled samples, improve multi-source image consistency, and enhance model generalization across regions and years.

Study Configuration

• Spatial Scale: Middle and lower Yangtze River Basin (Anhui, Jiangsu, Zhejiang, Jiangxi, Hubei, Hunan, and Shanghai), covering approximately 1,000,000 square kilometers. Mapping resolution is 10 meters.
• Temporal Scale: Multi-annual mapping for 2022, 2023, and 2024. Satellite data were acquired from April to November for each year. Temporal transferability was also evaluated for 2019-2021.

Methodology and Data

• Models used:
  • Progressive Deep Learning Crop Mapping (PDLCM) framework, an iterative enhancement-based learning strategy.
  • Spatio-temporal Fusion Deep Learning Model Based on Multi-Source Data (SFDLM-MSD), combining:
    • Long Short-Term Memory (LSTM) modules for temporal feature extraction from Synthetic Aperture Radar (SAR) time-series imagery.
    • U-Net modules for processing spatial information from optical images and fused temporal features.
  • Random Forest classifier for initial sample expansion (Spatial Discrete Samples Expansion - SDSE).
  • t-Distributed Stochastic Neighbor Embedding (t-SNE) for feature visualization.
  • Kernel density estimation and Moran’s I index for spatial distribution analysis.
• Data sources:
  • Satellite:
    • Sentinel-2 optical data (visible, near-infrared, and shortwave infrared bands B2-B8A, B11, B12), composited into 14-band seasonal images (10 spectral bands + temporal maximum/minimum Normalized Difference Vegetation Index (NDVI) and Land Surface Water Index (LSWI)).
    • Sentinel-1 SAR data (VV and VH polarizations), acquired at 12-day intervals, resulting in 36 radar composites.
  • Ground Truth/Validation:
    • Field surveys with geo-tagged photographs and farmer interviews across six provinces.
    • Approximately 18,000 randomly generated validation sample points (6000 per year for 2022-2024) labeled via visual interpretation using Google Earth, Esri World Imagery, multi-temporal satellite imagery, and field survey data.
    • Rice cultivation statistics from provincial and municipal statistical yearbooks and government reports.
    • Multi-temporal Unmanned Aerial Vehicle (UAV) monitoring for detailed phenological observations and distinguishing single/double-season rice.
  • Auxiliary Data for Transferability Analysis: Digital Elevation Model (DEM), annual precipitation, average temperature, climate types, soil types, and maturity periods of early and middle rice.
  • Platform: Google Earth Engine (GEE) for large-scale image filtering, cloud masking, temporal compositing, and batch downloading; local processing for deep learning model development, training, and prediction.

Main Results

• The PDLCM framework successfully generated 10 m resolution rice distribution maps for the middle and lower Yangtze River Basin for 2022, 2023, and 2024.
• For the Jiangsu, Zhejiang, and Shanghai (JZS) region in 2022, the model achieved an overall accuracy (OA) of 97.54%, an F1-score of 0.922, a Kappa coefficient of 0.908, a precision of 0.963, and a recall of 0.885.
• Across the entire study area for 2022-2024, the model demonstrated stable and high accuracy with an average OA of 96.81%, an F1-score of 0.882, a recall of 0.854, a precision of 0.913, and a Kappa coefficient of 0.864.
• Spatial transferability analysis showed a distance decay pattern, with the F1-score decreasing from 0.929 at 120 kilometers to 0.848 at 400 kilometers from the training area.
• Temporal transferability was generally good but varied regionally, with some areas experiencing reduced accuracy due to interannual variability, data quality issues (e.g., cloud cover), and local agricultural practices.
• Rice cultivation was found to be mainly distributed in flat terrain, fertile soils, and areas with abundant water, with high-density zones (mean densities > 41.6 hectares per square kilometer) concentrated in central/northern Jiangsu, central Anhui, central Hubei, northern Hunan, and central Jiangxi.
• A strong correlation (R² = 0.851) was observed between remote sensing-derived rice planting areas and official municipal statistical yearbook data.
• All generated rice mapping products (2022–2024) and source codes are publicly available.

Contributions

• Developed a novel Progressive Deep Learning Crop Mapping (PDLCM) framework that enables accurate national-scale 10 m rice mapping using sparse initial samples, effectively mitigating sample scarcity and spatial heterogeneity challenges.
• Integrated time-series Sentinel-1 SAR and Sentinel-2 optical data with a spatio-temporal fusion deep learning model (SFDLM-MSD) to robustly capture rice phenological characteristics and address data inconsistencies over large areas.
• Produced the first deep learning–based, multi-annual (2022-2024), high-resolution (10 m) rice distribution maps for the entire middle and lower Yangtze River Basin (over 1,000,000 square kilometers), making these valuable datasets publicly accessible.
• Demonstrated the strong spatial and temporal generalization capabilities of the framework, offering a scalable and transferable paradigm for operational crop mapping that reduces the reliance on extensive labeled datasets.
• Identified and analyzed key environmental and data-related factors (e.g., climatic/phenological gradients, terrain, soil, image availability/quality) influencing model transferability, providing critical insights for future model refinements and broader applicability.
• Provided reliable spatial information through the publicly available multi-year 10 m rice maps, supporting food security assessment, agricultural management, and sustainability-related policy making, with direct relevance to Sustainable Development Goal 2 (SDG 2) monitoring.

Funding

• National Natural Science Foundation of China (42201412, 42401400)
• National Key R&D Program of China (2023YFB3906201-02, 2022YFD2000100)

Citation

@article{Yang2026From,
  author = {Yang, Lingbo and Dong, Jiancong and Xu, Cong and Huang, Jingfeng and Wang, Yan and Lü, H. and Chen, Z. and Le, Wang and Zhang, Jingcheng},
  title = {From Sparse to Refined Samples: Iterative Enhancement-Based PDLCM for Multi-Annual 10 m Rice Mapping in the Middle-Lower Yangtze},
  journal = {Remote Sensing},
  year = {2026},
  doi = {10.3390/rs18020209},
  url = {https://doi.org/10.3390/rs18020209}
}