Hu et al. (2025) High-Accuracy Identification of Cropping Structure in Irrigation Districts Using Data Fusion and Machine Learning

Identification

• Journal: Remote Sensing
• Year: 2025
• Date: 2025-12-27
• Authors: Xinli Hu, Changming Cao, Ziyi Zan, Kun Wang, Meng Chai, Su Li, Weifeng Yue
• DOI: 10.3390/rs18010101

Research Groups

• College of Water Sciences, Beijing Normal University, Beijing, China
• College of Energy and Environment, Inner Mongolia University of Science and Technology, Baotou, China

Short Summary

This study developed a high spatiotemporal-resolution remote sensing approach for identifying cropping structures in heterogeneous irrigation districts by fusing Landsat, Sentinel-2, and MODIS data to create a continuous 30 m, 8-day Normalized Difference Vegetation Index (NDVI) time series. Utilizing phenology-based features and a Random Forest classifier, the method achieved an overall accuracy of 90.78% and a Cohen’s kappa coefficient of 0.882 for crop mapping in the Yichang Irrigation District.

Objective

• To develop a high spatiotemporal-resolution remote sensing approach for accurate crop identification and regional crop distribution mapping in heterogeneous irrigation districts, addressing challenges posed by persistent cloud cover and mosaic cropping patterns.

Study Configuration

• Spatial Scale: Yichang Irrigation District (YCID), Hetao Plain, China (approximately 3352.5 km²). Output crop distribution maps at 30 m spatial resolution.
• Temporal Scale: Growing seasons (April–October) of 2023 (training) and 2024 (validation). Reconstructed NDVI time series at 8-day temporal resolution.

Methodology and Data

• Models used:
  • Spatiotemporal Adaptive Reflectance Fusion Model (STARFM) for data fusion.
  • Random Forest (RF) classifier for crop discrimination.
  • SHapley Additive exPlanations (SHAP) for feature importance and interpretability.
  • Simple Non-Iterative Clustering (SNIC) for superpixel-level smoothing.
• Data sources:
  • Satellite: MODIS MOD09GA (Collection 6.1), Landsat surface reflectance (Collection 2, Level-2: Landsat-5 TM, Landsat-7 ETM+, Landsat-8/9 OLI), Sentinel-2 MSI Level-2A.
  • Ground-truth: In-situ field sample points collected in 2023 (2508 samples for training) and 2024 (312 samples for independent validation) using Ovitalmap mobile mapping application with GNSS.
  • Ancillary: ERA5 for climatic data (2 m air temperature, total precipitation).
  • Platform: Google Earth Engine (GEE).

Main Results

• A continuous 30 m, 8-day NDVI time series was successfully reconstructed through tri-source fusion of Landsat, Sentinel-2, and MODIS, significantly reducing cloud-induced gaps and enhancing spatial detail compared to single-sensor or low-resolution products.
• Class-specific NDVI trajectories revealed distinct phenological patterns for wheat (early, short peak in June), maize (extended high-value plateau from July to August), and sunflower (latest onset, peak in August), providing strong separability.
• SHAP analysis identified late-June to early-July NDVI values (NDVI0620, NDVI0628, NDVI0706, NDVIm06) as the most dominant features for crop class separability.
• The Random Forest classifier, utilizing 11 phenological features derived from the fused NDVI series, achieved an overall accuracy (OA) of 90.78% and a Cohen’s kappa (κ) coefficient of 0.882 for the 2024 crop distribution map.
• Per-class accuracy metrics showed high F1-scores (close to 1.0) for non-crop categories (building, water, forest), while maize and sunflower exhibited producer accuracies of approximately 0.80 and 0.70, respectively, indicating some residual confusion between these summer crops.
• Areal statistics for the Yichang Irrigation District in 2024 indicated that sunflower and maize were the dominant crops, covering 44.4% and 33.2% of the study area, respectively, with wheat occupying 2.2%.

Contributions

• Developed and validated a novel tri-source (Landsat, Sentinel-2, MODIS) spatiotemporal data fusion framework to reconstruct a high-resolution (30 m, 8-day) NDVI time series, demonstrating improved temporal continuity, reduced cloud noise, and enhanced preservation of parcel boundaries and phenological information.
• Designed a phenology-based feature set, combining anchor-date NDVI values with curve-shape descriptors, which significantly improved fine-grained crop discrimination compared to methods lacking fusion or explicit phenological constraints.
• Integrated SHapley Additive exPlanations (SHAP) to provide interpretable insights into feature contributions and their class-discriminative effects, guiding feature optimization and confirming the reliance on phenology-level contrasts.
• Provided a robust, accurate, and transferable framework for regional-scale precision agricultural monitoring and cropping structure extraction in complex, heterogeneous irrigated environments.

Funding

• National Natural Science Foundation of China (Grant No. 52179032 & U24A20179).

Citation

@article{Hu2025HighAccuracy,
  author = {Hu, Xinli and Cao, Changming and Zan, Ziyi and Wang, Kun and Chai, Meng and Li, Su and Yue, Weifeng},
  title = {High-Accuracy Identification of Cropping Structure in Irrigation Districts Using Data Fusion and Machine Learning},
  journal = {Remote Sensing},
  year = {2025},
  doi = {10.3390/rs18010101},
  url = {https://doi.org/10.3390/rs18010101}
}