County-Level Soybean Yield Prediction Using Deep CNN-LSTM Model | JoVE Visualize

Area of Science:

Agricultural Science
Remote Sensing
Machine Learning
Deep Learning

Background:

Accurate crop yield prediction is crucial for agricultural management, market planning, and insurance.
Remote sensing data combined with machine learning, particularly Deep Learning (DL), has shown promise in yield prediction.
Convolutional Neural Networks (CNNs) excel at spatial feature extraction, while Long Short-Term Memory (LSTM) networks capture temporal dependencies in crop phenology.

Purpose of the Study:

To propose and evaluate a novel deep Convolutional Neural Network-Long Short-Term Memory (CNN-LSTM) model for soybean yield prediction.
To assess the model's performance for both end-of-season and in-season yield predictions at the county level across the continental United States (CONUS).
To investigate the potential of integrating spatial and temporal features for enhanced yield prediction accuracy.

Main Methods:

Developed a hybrid CNN-LSTM deep learning model trained on crop growth and environmental variables.
Utilized Google Earth Engine (GEE) to combine and process remote sensing data (MODIS Land Surface Temperature and Surface Reflectance) and weather data.
Transformed combined data into histogram-based tensors for deep learning input, using historical soybean yield data as labels.

Main Results:

The proposed CNN-LSTM model demonstrated superior prediction performance compared to standalone CNN or LSTM models.
The hybrid model achieved higher accuracy in both end-of-season and in-season soybean yield predictions.
The methodology shows effectiveness in county-level yield prediction using integrated remote sensing and environmental data.

Conclusions:

The CNN-LSTM model offers a significant advancement in crop yield prediction accuracy by effectively leveraging both spatial and temporal data.
This approach holds substantial potential for improving yield prediction for other major crops like corn, wheat, and potatoes at fine scales.
The study highlights the value of combining different deep learning architectures for complex agricultural prediction tasks.