Pre-season crop type mapping using deep neural networks

Reliable crop type maps are needed early in the growing season to retrieve crop type information from satellite imagery in order to produce in-season crop condition and yield outlooks. However, the inability to have early crop maps due to limited earth observations available during the beginning of the season is a major challenge for developing reliable satellite-based early warning systems. This paper introduces a novel crop type prediction modeling system based on deep Neural Networks (NN) to produce preseason crop type maps at the field scale resolution using historical crop maps. The architecture of this modeling system comprises of two end-to-end NN based modules that form an autoencoder configuration: a spatio-temporal Encoder, built off of the Bidirectional ConvLSTM network, and a Decoder, which can learn both spatial and temporal patterns necessary to accurately predict the crop sequences. To build this system, we used USDA-NASS historical Cropland Data Layer (CDL) data of Nebraska and trained nine different neural network models using a combination of three rotation cycles (3-, 4- and 5-year rotations) and three sets of historical CDL data with varying durations (2010-2016; 2006-2016; and 2002-2016). All trained models performed well when applied to predict 2017 land cover, and achieved maximum accuracy of 88%. However, the amount of training time taken to reach 88% accuracy varies for each model. Models with more temporal information were found to learn faster and achieve maximum accuracy quickly. Further, we compared the NN model with a Markov Chain (MC) based approach, a common approach used for crop pattern recognition, by applying them to predict 2018 and 2019 land cover maps using 2006-2016 CDL data. Results showed that the NN model outperformed the MC model with higher overall accuracy (0.77) and kappa coefficient (0.57) compared to that of the MC model (overall accuracy = 0.67 and kappa coefficient = 0.44). In summary, our novel system based on deep neural networks showed that it can learn complex spatial and temporal patterns from the historical land cover data and produce reasonable early crop maps, which can be used in satellite-based early warning systems.