Improving wheat yield estimates using data augmentation models and remotely sensed biophysical indices within deep neural networks in the Guanzhong Plain, PR China

Crop yield estimation and prediction constitutes a key issue in agricultural management, particularly under the context of demographic pressure climate change. Currently, main challenge estimating crop yields based on remotely sensed data data-driven methods is how to cope with small datasets limited amount annotated samples. In order samples improve accuracy winter wheat Guanzhong Plain, PR China, this study proposed method combining generative adversarial networks (GANs) convolutional neural network (CNN) for comprehensive growth monitoring wheat, which leaf area index (LAI), vegetation temperature condition (VTCI) meteorological at four stages during 2012–2017 were generated as inputs multi-layer (CNNs), GAN was employed artificially increase number training Then, linear regression model between simulated (I) measured established estimate Plain pixel by pixel. The final results showed when used double size samples, simulation values obtained CNN augmented using provided better (R2 = 0.95, RMSE 0.05), validation 0.54, 0.16) testing 0.50, 0.14) performance than that just original achieved best pixel-scale 591.46 kg/ha) Plain. These can be enlarged GAN, thus, more important features reflecting conditions from indices extracted, indicated accompanied could contribute lot augmentation are extremely useful application deep learning.