Deep Hybrid Neural Networks on Multi-temporal Satellite Data: Predicting Farm-scale Crop Yields

Abstract

Accurate farm-scale crop yield predictions can enable farmers to improve their yield per decare and inform subsequent sectors of the availability of grains sooner. Existing research on yield predictions is limited to regional analytics, which often fails to capture local yield variations influenced by farm management decisions and field conditions. Farm-scale crop yield predictions require precise ground-truth prediction targets, which are not always available. It takes substantial manual labor to create large and suitable datasets of high-resolution per-farm samples. This thesis introduces a hybrid multi-temporal deep neural network that combines convolutional and recurrent features specially designed to predict the individual crop yields of farms across Norway with per-farm samples. To the best of our knowledge, this is the first farm-scale crop yield prediction model of its kind. The hybrid model learns to extract features from both multi-temporal satellite images and weather data time series to predict crop yields accurately. We use a complex multitude of noisy data sources, including multi-temporal satellite images from Sentinel-2, weather data from The Norwegian Meteorological Institute, farm data and grain delivery data from the Norwegian Agriculture Agency, and cadastral data. Our hybrid model, which combines two and one-dimensional convolutional layers and a gated recurrent unit network, predicts crop yields with an error of 76 kg/daa using satellite images and weather data, according to our experiments.