Predicting wheat yield using deep learning and multi-source environmental data.

Area of Science:

• Agricultural Science
• Data Science
• Environmental Science

Background:

• Accurate crop yield forecasting is vital for food security and sustainable agriculture.
• Winter wheat yield prediction in Pakistan is complex due to interacting climatic, soil, and environmental factors.

Purpose of the Study:

• To introduce DeepAgroNet, a novel deep learning framework for estimating winter wheat yields in southern Pakistan.
• To integrate satellite imagery, meteorological data, and soil characteristics for improved yield prediction.

Main Methods:

• Developed a three-branch deep learning framework (DeepAgroNet) using convolutional neural networks (CNN), recurrent neural networks (RNN), and artificial neural networks (ANN).
• Trained models on detrended winter wheat yield data from 2017-2022.
• Utilized Google Earth Engine for processing remote sensing, climate, and soil data.

Main Results:

• The CNN model achieved the highest accuracy (R²=0.77, 98% forecast accuracy one month pre-harvest).
• RNN and ANN models showed moderate predictive capabilities (R²=0.72 and R²=0.66, respectively).
• All models demonstrated yield error rates below 10%, effectively integrating diverse data types.

Conclusions:

• DeepAgroNet effectively integrates spatial, temporal, and static data for reliable winter wheat yield prediction.
• The framework offers a scalable solution for precision agriculture, enhancing food security and sustainable development in Pakistan.
• The adaptable DeepAgroNet framework can be applied to other agricultural regions globally.