Is deeper always better? Evaluating deep learning models for yield forecasting with small data.

Area of Science:

• Agricultural Science
• Data Science
• Machine Learning

Background:

• Accurate crop yield prediction is crucial for food security, particularly in vulnerable nations.
• Forecasting anomalously low yields requires robust and adaptable modeling techniques.
• Satellite data offers a globally accessible resource for near real-time agricultural monitoring.

Purpose of the Study:

• To investigate a flexible deep learning approach for provincial-level crop yield forecasting.
• To assess the performance of 1D and 2D convolutional neural networks (CNNs) with limited data.
• To compare deep learning models against conventional benchmarks and machine learning algorithms.

Main Methods:

• Utilized deep 1D and 2D CNNs with hyperparameter optimization for model selection.
• Input data included 3D histograms of Normalized Difference Vegetation Index (NDVI) and climate data for the 2D CNN.
• Employed time series averages of NDVI and climate data for the 1D CNN, applied to Algerian crop data (2002-2018).

Main Results:

• Deep learning models demonstrated suboptimal performance compared to traditional benchmarks and machine learning algorithms.
• Simple benchmarks like peak NDVI were difficult to surpass.
• Machine learning models consistently outperformed deep learning across all crops and forecasting months.

Conclusions:

• The limited size of the available dataset was identified as the primary reason for the deep learning models' poor performance.
• The developed deep learning approach, while flexible and transferable, requires larger datasets for effective yield prediction.
• Further research should focus on data augmentation or transfer learning strategies to improve deep learning efficacy with sparse data.