Batkalova et al. (2026) Training Sample Migration for Temporal Cropland Mapping in Central Asia

Identification

• Journal: Land
• Year: 2026
• Date: 2026-01-13
• Authors: Aiman Batkalova, Pengyu Hao
• DOI: 10.3390/land15010156

Research Groups

• Digital FAO and Agro-Informatics Division, Food and Agriculture Organization of the United Nations, Rome, Italy

Short Summary

This study developed an NDVI-based training sample migration framework for temporal cropland mapping in Central Asia, transferring labeled samples from reference years to a target year using time-series similarity analysis. The framework achieved high overall accuracies of 86% in Kazakhstan and 95% in Uzbekistan, demonstrating improved spatial coherence and reduced misclassification compared to global cropland products.

Objective

• To propose and evaluate a tailored NDVI-based training sample migration framework that systematically assesses multiple time-series similarity metrics and their combinations to identify temporally stable cropland samples.
• To use these migrated samples to train a Random Forest classifier for annual cropland mapping in data-scarce and climatically challenging environments of Central Asia.

Study Configuration

• Spatial Scale: Two pilot regions in Central Asia: Akmola Region (northern Kazakhstan, rainfed wheat systems) and Kashkadarya Region (southern Uzbekistan, irrigated cotton-dominated systems).
• Temporal Scale: Biweekly Normalized Difference Vegetation Index (NDVI) time series (24 layers per year) for reference years (2016 for Uzbekistan, 2022 for Kazakhstan) and the target year (2021).

Methodology and Data

• Models used: Random Forest (RF) classifier.
• Data sources:
  • Satellite Imagery: Landsat-8 (LANDSAT/LC08/C02/T1L2) and Sentinel-2 (COPERNICUS/S2SR) surface reflectance products for biweekly NDVI composites.
  • Reference Samples:
    • Kazakhstan (2022): AgroOpen national satellite monitoring platform (Joint Stock Company “Kazakhstan Gharysh Sapary”).
    • Uzbekistan (2016): ESA WorldCereal project reference data.
  • Validation Samples: Stratified random samples for 2021, visually interpreted using Sentinel-2 10 m imagery, NDVI time series profiles, Google Earth, Bing Maps, and FAO Collect Earth platform.
  • Similarity Metrics: Ten NDVI-based time-series similarity metrics were evaluated: Euclidean distance, Chebyshev distance, Weighted Minkowski distance, Cubic distance, Dynamic Time Warping (DTW), Spectral Angle Distance (SAD), Pearson correlation, Spearman correlation, Kendall Tau, and Cosine similarity.
  • Comparison Products: ESA WorldCereal 2021 and ESA WorldCover 2021 global cropland products.

Main Results

• The proposed migration framework achieved high overall accuracies of 86% in the Akmola Region (Kazakhstan) and 95% in the Kashkadarya Region (Uzbekistan) for 2021 cropland mapping, based on independent validation.
• Hybrid combinations of similarity metrics consistently outperformed single-metric approaches, improving overall classification accuracy by 5–8% due to enhanced robustness to phenological variability and interannual differences.
• Comparison with global cropland products (WorldCereal 2021 and WorldCover 2021) showed that the migration-based maps provided improved spatial coherence, sharper field boundaries, and reduced misclassification, particularly in semi-arid transition zones where global products tended to overestimate cropland.
• In Kazakhstan, the migration-based map achieved an overall accuracy of 0.93 (vs. WorldCereal 2021) and 0.87 (vs. WorldCover 2021), with high precision (0.94) and recall (0.96) for non-cropland, indicating effective reduction of commission errors.
• In Uzbekistan, the migration-based map achieved an overall accuracy of 0.92 (vs. WorldCereal 2021) and 0.88 (vs. WorldCover 2021), demonstrating sharper cropland boundaries and more accurate exclusion of non-agricultural areas.

Contributions

• Introduces a systematic and validated NDVI time-series-based training sample migration framework for annual cropland mapping, extending beyond single-metric or heuristic strategies.
• Demonstrates the effectiveness of reusing high-quality reference samples across years, reducing the reliance on repeated annual field surveys in data-scarce regions.
• Provides a comprehensive evaluation and integration of ten diverse similarity metrics (geometric, temporal alignment, and correlation-based), showing that combining complementary metrics enhances robustness and transferability in heterogeneous and drought-prone agricultural landscapes.
• Achieves improved spatial coherence, sharper field boundaries, and reduced misclassification in semi-arid transition zones compared to existing global cropland products (WorldCereal and WorldCover).
• Offers a cost-effective and repeatable pathway for regional cropland monitoring under contrasting agricultural practices and climatic conditions, particularly in Central Asia.

Funding

This research received no external funding.

Citation

@article{Batkalova2026Training,
  author = {Batkalova, Aiman and Hao, Pengyu},
  title = {Training Sample Migration for Temporal Cropland Mapping in Central Asia},
  journal = {Land},
  year = {2026},
  doi = {10.3390/land15010156},
  url = {https://doi.org/10.3390/land15010156}
}