Zhang et al. (2025) Long history paddy rice mapping across Northeast China with deep learning and annualresult enhancement method

Identification

• Journal: Earth system science data
• Year: 2025
• Date: 2025-12-05
• Authors: Zihui Zhang, Lang Xia, Fen Zhao, Yue Gu, Jing Yang, Yan Zha, Shangrong Wu, Peng Yang
• DOI: 10.5194/essd-17-6851-2025

Research Groups

• State Key Laboratory of Efficient Utilization of Arid and Semi-arid Arable Land in Northern China, the Institute of Agricultural Resources and Regional Planning, Chinese Academy of Agricultural Sciences, Beijing, China
• Key Laboratory of Agricultural Remote Sensing, Ministry of Agriculture and Rural Affairs, Beijing, China
• Department of Electrical and Computer Engineering, Rutgers University, New Brunswick, USA

Short Summary

This study developed a deep learning and annual result enhancement (ARE) method to generate annual paddy rice maps for Northeast China from 1985 to 2023 using multi-sensor Landsat data, demonstrating significantly improved accuracy over traditional methods. The research revealed a substantial expansion of paddy rice cultivation in the region, providing valuable data for agricultural monitoring and policy.

Objective

• To generate annual paddy rice maps with 30 meter spatial resolution for Northeast China from 1985 to 2023 using multi-sensor Landsat data and a deep learning model.
• To develop an Annual Result Enhancement (ARE) method to improve mapping accuracy by mitigating the impact of limited training samples and cross-sensor differences across various phenological stages.
• To construct a cross-sensor training dataset for paddy rice using Landsat 5 TM and Landsat 8 OLI sensors.

Study Configuration

• Spatial Scale: Northeast China (Heilongjiang, Jilin, Liaoning provinces, and northeastern Inner Mongolia), covering approximately 1.26 x 10^6 square kilometers.
• Temporal Scale: Annual mapping from 1985 to 2023 (39 years).

Methodology and Data

• Models used:
  • Full-Resolution Network (FR-Net): A deep learning semantic segmentation model.
  • Annual Result Enhancement (ARE) method: Developed to integrate category probabilities and confidence levels from FR-Net across phenological stages.
  • XGBoost classifier: Used for initial generation of cross-sensor training data.
• Data sources:
  • Satellite Imagery: 13,809 Landsat Collection 2 Level-2 surface reflectance products (Landsat 5 Thematic Mapper, Landsat 8/9 Operational Land Imager) with 30 meter spatial resolution, collected during May to September (1985-2023).
  • Training Dataset: A cross-sensor paddy training dataset comprising 155 Landsat scenes, generated using XGBoost and manual visual correction.
  • Ground Truth Data: 107,954 ground truth samples collected from field surveys (2011-2023) and Google Earth Very High Resolution (VHR) imagery (2002-2023).
  • Agricultural Statistics Data: Paddy planted areas from district, municipal, and provincial statistical bureaus (1985-2022).

Main Results

• The Annual Result Enhancement (ARE) method significantly improved mapping accuracy, showing a 6% increase in Overall Accuracy (OA), a 5% increase in F1 score, and a 13% increase in Matthews Correlation Coefficient (MCC) compared to traditional rice mapping methods.
• The overall mapping results achieved high average accuracy metrics: User Accuracy (UA) of paddy = 0.93, Producer Accuracy (PA) of paddy = 0.91, Overall Accuracy (OA) = 0.91, F1 score = 0.92, and MCC = 0.82.
• The paddy rice cultivation area in Northeast China increased from 1.11 x 10^4 square kilometers in 1985 to 6.45 x 10^4 square kilometers in 2023, representing an overall expansion of 5.34 x 10^4 square kilometers (a 4.81-fold increase).
• Heilongjiang province experienced the largest increase in paddy area, with a gain of 4.33 x 10^4 square kilometers between 1985 and 2023.
• The generated paddy rice maps showed strong consistency and correlation with agricultural statistics data, with R^2 values of 0.93 for the entire study area and 0.96 at provincial, municipal, and district levels.

Contributions

• Constructed a novel cross-sensor training dataset for paddy rice using Landsat 5 TM and Landsat 8 OLI sensors, enabling long-term mapping across different sensor generations.
• Proposed an innovative Annual Result Enhancement (ARE) method that leverages category probabilities from a deep learning model across different phenological stages to significantly improve annual paddy rice mapping accuracy, particularly under conditions of limited training samples.
• Generated the first yearly paddy rice maps with 30 meter spatial resolution for Northeast China spanning from 1985 to 2023, providing a consistent and long-history dataset for monitoring paddy rice dynamics and assessing land-use policies.

Funding

• National Key Research and Development Program of China (grant no. 2022YFD2001105)
• National Natural Science Foundation of China (grant no. 42401448/42301442)
• Fundamental Research Funds for Central Non-profit Scientific Institution (grant no. 1610132025003)
• Open project of State Key Laboratory of Efficient Utilization of Arid and Semi-arid Arable Land in Northern China, the Institute of Agricultural Resources and Regional Planning, Chinese Academy of Agricultural Sciences (grant no. EUAL-2023-02)
• Agricultural Science and Technology Innovation Project of the Chinese Academy of Agriculture Sciences

Citation

@article{Zhang2025Long,
  author = {Zhang, Zihui and Xia, Lang and Zhao, Fen and Gu, Yue and Yang, Jing and Zha, Yan and Wu, Shangrong and Yang, Peng},
  title = {Long history paddy rice mapping across Northeast China with deep learning and annualresult enhancement method},
  journal = {Earth system science data},
  year = {2025},
  doi = {10.5194/essd-17-6851-2025},
  url = {https://doi.org/10.5194/essd-17-6851-2025}
}