Peng et al. (2025) Research on dynamic prediction of vegetation coverage by precipitation-evapotranspiration in arid regions based on CNN-LSTM hybrid model

Identification

• Journal: Scientific Reports
• Year: 2025
• Date: 2025-12-17
• Authors: Kai Peng, Yanfei Zhang, Tiejun Liu, Zijing Li, Wentao Liang, Hualin Liu, Yawen Bai
• DOI: 10.1038/s41598-025-31530-z

Research Groups

• Yinshanbeilu Grassland Eco-hydrology National Observation and Research Station, China Institute of Water Resources and Hydropower Research, Beijing 100038, PR China
• Institute of Pastoral Hydraulic Research, MWR, Hohhot, 010020, China
• Collaborative Innovation Center for Grassland Ecological Security (Jointly Supported by the Ministry of Education of China and Inner Mongolia Autonomous Region), Hohhot, 010021, China

Short Summary

This study developed a CNN-LSTM hybrid model to dynamically predict vegetation coverage in arid regions, integrating SPEI-based drought classification and precipitation-evapotranspiration data. The model achieved a Pearson correlation coefficient of 0.95 with measured data, accurately capturing vegetation dynamics from 2000 to 2022.

Objective

• To develop an accurate and effective CNN-LSTM hybrid model for dynamic prediction of vegetation coverage in arid regions, considering phenological lag and the dominant influence of precipitation and evapotranspiration, and to quantify vegetation coverage changes using dynamic change indices.

Study Configuration

• Spatial Scale: Inner Mongolian Gobi region, including Ejina Banner, Wulate Rear Banner, Wulate Middle Banner, and adjacent zones, extending to Alxa League, China-Mongolia border, Bayannur City, and Gansu Province's Hexi Corridor.
• Temporal Scale: 2000 to 2022 (23 years).

Methodology and Data

• Models used:
  • CNN-LSTM hybrid model for dynamic vegetation coverage prediction.
  • Standardized Precipitation Evapotranspiration Index (SPEI) for drought classification.
  • Hamon method for Potential Evapotranspiration (PET) calculation.
  • Three-parameter log-logistic probability distribution function for cumulative moisture deficit series.
  • Kriging interpolation method for spatialization of meteorological station data.
  • Convolutional Neural Network (CNN) for spatial feature extraction.
  • Long Short-Term Memory (LSTM) network for capturing temporal dependencies.
  • Adam optimizer, ReLU activation function, Softmax function, Sigmoid function, Tanh function.
  • Benchmark models for comparison: Multiple linear regression, Support Vector Machine (SVM) regression, single CNN, single LSTM.
• Data sources:
  • Annual precipitation data: China Meteorological Science Data Sharing Service Network (http://cdc.cma.gov.cn/).
  • Annual evaporation (potential evapotranspiration) data: Resource and Environment Data Cloud Platform of the Chinese Academy of Sciences (http://www.resdc.cn/).
  • Daily data from 118 meteorological stations in Inner Mongolia Autonomous Region and surrounding areas.
  • Vegetation coverage data: Resource and Environment Data Cloud Platform of the Chinese Academy of Sciences (http://www.resdc.cn/), derived from MODIS remote sensing images.
  • Land use data: Resource and Environment Data Cloud Platform of the Chinese Academy of Sciences.

Main Results

• The CNN-LSTM hybrid model achieved a Pearson correlation coefficient of 0.95 and a Root Mean Square Error (RMSE) of 0.042 on the test set, demonstrating superior performance compared to benchmark models (e.g., single LSTM R=0.91, RMSE=0.065; single CNN R=0.88, RMSE=0.071).
• The model effectively identified differences in rainfall characteristics across drought severity levels. In arid regions, the contribution rate of heavy rainfall to the total rainfall anomaly (25.15%) was substantially higher than in semi-arid regions (1.55%), with statistical tests confirming a significant difference (P < 0.05).
• Potential evapotranspiration exhibited distinct spatial differentiation in the Inner Mongolian Gobi, generally increasing from the northeast (<700 mm/year) to the southwest (>1400 mm/year).
• The dynamic change index system (single-year and comprehensive) accurately captured phased variations in vegetation coverage from 2000 to 2022, showing an overall upward trend despite fluctuations. Both indices had a 0.95 correlation coefficient with observed dynamics.
• Vegetation coverage grades correlated with average coverage: lower average coverage (e.g., 2000-2006) corresponded to predominantly medium-to-low grades, while higher average coverage (e.g., 2010-2012, 2016-2022) corresponded to mostly medium-to-high and high grades.

Contributions

• Developed an innovative CNN-LSTM hybrid modeling approach for dynamic vegetation coverage prediction in arid regions, integrating SPEI-based drought classification and effectively addressing phenological lag and multi-factor interactions.
• Introduced a novel dynamic change index system (single-year and comprehensive) for precise quantification and comprehensive understanding of both short-term interannual fluctuations and long-term trends in vegetation coverage.
• Validated the effectiveness and transferability of the CNN-LSTM framework for "climate-vegetation" coupling prediction, applicable to other regions or similar ecological prediction tasks (e.g., soil moisture, net primary productivity).
• Revealed that high-intensity, low-frequency rainfall events are key hydrological drivers of vegetation changes in extremely arid regions, emphasizing the critical role of understanding pulse precipitation effects in the context of future climate change.

Funding

• China Institute of Water Resources and Hydropower Research Basic Research Business Fund Special Project (Grant No. MKSZ2024JK014)
• National Natural Science Foundation of China (U24A20573)

Citation

@article{Peng2025Research,
  author = {Peng, Kai and Zhang, Yanfei and Liu, Tiejun and Li, Zijing and Liang, Wentao and Liu, Hualin and Bai, Yawen},
  title = {Research on dynamic prediction of vegetation coverage by precipitation-evapotranspiration in arid regions based on CNN-LSTM hybrid model},
  journal = {Scientific Reports},
  year = {2025},
  doi = {10.1038/s41598-025-31530-z},
  url = {https://doi.org/10.1038/s41598-025-31530-z}
}