Assessment of Crop Yield Prediction Capabilities of CNN Using Multisource Data

Abstract

The growing abundance of digitally available spatial, geological, and climatological data opens up new opportunities for agricultural data-based input–output modeling. In our study, we took a convolutional neural network model previously developed on Unmanned Aerial Vehicle (UAV) image data only and set out to see whether additional inputs from multiple sources would improve the performance of the model. Using the model developed in a preceding study, we fed field-specific data from the following sources: near-infrared data from UAV overflights, Sentinel-2 multispectral data, weather data from locally installed Vantage Pro weather stations, topographical maps from National Land Survey of Finland, soil samplings, and soil conductivity data gathered with a Veris MSP3 soil conductivity probe. Either directly added or encoded as additional layers to the input data, we concluded that additional data helps the spatial point-in-time model learn better features, producing better fit models in the task of yield prediction. With data of four fields, the most significant performance improvements came from using all input data sources. We point out, however, that combining data of various spatial or temporal resolution (i.e., weather data, soil data, and weekly acquired images, for example) might cause data leakage between the training and testing data sets when training the CNNs and, therefore, the improvement rate of adding additional data layers should be interpreted with caution.