Wheat yield estimation using remote sensing data based on machine learning approaches.

Area of Science:

• Agricultural Science
• Remote Sensing
• Data Science

Background:

• Accurate wheat yield prediction is crucial for agricultural planning and global trade.
• Traditional linear regression models provide insufficient accuracy for Chinese wheat crop productivity.
• Advanced data-driven approaches are needed to enhance yield estimation.

Purpose of the Study:

• To compare the performance of four data-driven models (LSTM, RF, GBDT, SVR) for winter wheat yield prediction.
• To evaluate the accuracy of yield estimates using Sentinel-2 (multispectral) and ZY-1 02D (hyperspectral) data.
• To identify the most effective method and data source for precise winter wheat yield forecasting.

Main Methods:

• Utilized Sentinel-2 multispectral and ZY-1 02D hyperspectral data.
• Employed four machine learning algorithms: Long Short-Term Memory (LSTM), Random Forest (RF), Gradient Boosting Decision Tree (GBDT), and Support Vector Regression (SVR).
• Integrated 15,709 gridded yield data points (5m x 5m resolution) for model training and validation.

Main Results:

• The Long Short-Term Memory (LSTM) model achieved the highest accuracy with a Root Mean Square Error (RMSE) of 0.201 t/ha.
• Hyperspectral data from ZY-1 02D (RMSE = 0.237 t/ha) and 10m Sentinel-2 data (RMSE = 0.219 t/ha) outperformed 30m Sentinel-2 data (RMSE = 0.307 t/ha).
• The greenness vegetation index SR (simple ratio index) demonstrated superior performance over traditional vegetation indices.

Conclusions:

• Deep learning methods, particularly LSTM, offer significant advantages over traditional machine learning for accurate winter wheat yield prediction.
• Hyperspectral and higher-resolution multispectral data provide more reliable yield estimates than lower-resolution multispectral data.
• Shortwave infrared bands show potential for replacing visible and near-infrared bands in crop yield prediction models.