Li et al. (2025) Research on the estimation method of crop net primary productivity based on improved CASA model

Identification

• Journal: Frontiers in Plant Science
• Year: 2025
• Date: 2025-11-03
• Authors: Wanning Li, Zhuo Wang, Chunling Chen, Ying Ying Yin, Yuanji Cai, Hao Han, Minghuan Liu, Ziyi Feng
• DOI: 10.3389/fpls.2025.1659047

Research Groups

• College of Information and Electrical Engineering, Shenyang Agricultural University, Shenyang, China
• Henan Provincial Key Lab of Hydrosphere and Watershed Water Security, North China University of Water Resources and Electric Power, Zhengzhou, China

Short Summary

This research refines crop Net Primary Productivity (NPP) estimation by improving Fraction of Photosynthetically Active Radiation (FPAR) retrieval within the CASA model using a Convolutional Neural Network, significantly reducing FPAR Root Mean Square Error (RMSE) from 0.2040 to 0.0020 and NPP Mean Absolute Percentage Error (MAPE) from 28.92% to 20.31%.

Objective

• To develop a deep learning model that enhances the accuracy of FPAR estimation using high-resolution satellite imagery and vegetation indices from different crop growth stages.
• To integrate the deep learning framework with the CASA model for evaluating NPP estimation accuracy and validating the enhanced model’s reliability.
• To analyze the characteristics and driving factors of NPP results for different crops.

Study Configuration

• Spatial Scale: Haicheng City, Liaoning Province, China (122°39′18″E, 40°58′58″N), specifically the Shenyang Agricultural University experimental base in Gengzhuang Town.
• Temporal Scale: April to October 2022 for remote sensing data and meteorological data; field measurements conducted in 2022.

Methodology and Data

• Models used:
  • Carnegie-Ames-Stanford Approach (CASA) model (original and improved versions)
  • Convolutional Neural Network (CNN) for FPAR estimation
  • Recursive Feature Elimination (RFE) algorithm for feature selection
  • Gradient Boosted Decision Tree (GBDT) and Extreme Gradient Boosting (XGBoost) for comparison
  • Kriging interpolation for meteorological data
• Data sources:
  • Remote Sensing: High-resolution Sentinel-2 Level-2A satellite imagery (10 meter spatial resolution, 13 spectral bands, 5-10 day revisit period) covering April to October 2022.
  • Observation/Field: Field-measured Net Primary Productivity (NPP) data from 12 experimental plots (5 rice, 7 corn) in 2022, including plant dry weight and carbon content.
  • Meteorological: Monthly temperature (°C), precipitation (mm), and solar radiation (MJ/m²/month) for 2022 from the Xiaomaiya Agricultural Meteorological Big Data System platform.
  • Derived Data: 15 vegetation indices (ARVI, DVI, EVI, GEMI, GNDVI, MSAVI, NDI45, NDVI, PVI, RVI, REIP, SAVI, TNDVI, TSAVI, WDVI) calculated from Sentinel-2 imagery. FPAR data at 10 meter resolution generated by the Sentinel Application Platform (SNAP) biophysical processor (neural network trained on a radiative transfer model synthetic database) used as reference for CNN training.

Main Results

• The CNN model significantly improved FPAR estimation accuracy, reducing the average RMSE from 0.2040 (original CASA model's FPAR) to 0.0020 and the average MAE from 0.1984 to 0.0250.
• The Recursive Feature Elimination (RFE) algorithm consistently identified DVI, GEMI, NDI45, and RVI as key features for FPAR estimation.
• The CNN model outperformed GBDT and XGBoost for FPAR estimation, achieving an R² of 0.98, MSE of 0.0003, and MAE of 0.0127.
• The improved CASA model reduced the Mean Absolute Percentage Error (MAPE) for NPP estimation from 28.92% (original CASA) to 20.31%, an 8.61% improvement.
• Estimated NPP values in the study area ranged from 237.2 gC/m²/year to 891.1 gC/m²/year, with an average of 535.3 gC/m²/year.
• The improved model reduced the average absolute error for corn NPP estimates by 29% and for rice NPP estimates by 5.79%.
• Crop NPP peaked during June, July, August, and September, correlating with favorable meteorological conditions.

Contributions

• Developed a novel deep learning (CNN)-based framework for highly accurate FPAR estimation using high-resolution Sentinel-2 imagery and a comprehensive set of vegetation indices.
• Integrated the CNN-derived FPAR into the CASA model, significantly enhancing the precision and reliability of crop Net Primary Productivity (NPP) estimation compared to traditional methods.
• Demonstrated the superior performance of the CNN model for FPAR retrieval over statistical relationships and other machine learning algorithms (GBDT, XGBoost).
• Provided a robust methodology for large-scale crop NPP monitoring, offering strong support for agricultural decision-making, yield optimization, and carbon cycle studies.
• Addressed limitations of traditional FPAR estimation, such as saturation effects in high vegetation coverage, by leveraging CNN's ability to capture complex non-linear relationships.

Funding

• Liaoning Provincial Department of Education; Grant Number: JYTQN2023301.

Citation

@article{Li2025Research,
  author = {Li, Wanning and Wang, Zhuo and Chen, Chunling and Yin, Ying Ying and Cai, Yuanji and Han, Hao and Liu, Minghuan and Feng, Ziyi},
  title = {Research on the estimation method of crop net primary productivity based on improved CASA model},
  journal = {Frontiers in Plant Science},
  year = {2025},
  doi = {10.3389/fpls.2025.1659047},
  url = {https://doi.org/10.3389/fpls.2025.1659047}
}