A Bayesian Network Approach to Wheat Yield Prediction Using Topographic, Soil and Historical Data
1M. Karampoiki, 2L. Todman, 2S. Mahmood, 2A. Murdoch, 3J. Hammond, 1E. Ranieri, 1D. Paraforos
1. University of Hohenheim, Institute of Agricultural Engineering, Technology in Crop Production, Garbenstr. 9, 70599, Stuttgart, Germany
2. University of Reading, School of Agriculture, Policy and Development, Earley Gate, Reading RG6 6AR, United Kingdom
3. University of Reading, School of Agriculture, Policy and Development, Early Gate, Reading RG6 6AR, United Kingdom
4. Agricolus SRL, Via Settevalli 320, 06129 Perugia, Italy
Bayesian Network (BN) is the most popular approach for modeling in the agricultural domain. Many successful applications have been reported for crop yield prediction, weed infestation, and crop diseases. BN uses probabilistic relationships between variables of interest and in combination with statistical techniques the data modeling has many advantages. The main advantages are that the relationships between variables can be learned using the model as well as the potential to deal with missing data in some data entries. However, it is crucial to reduce data overfitting by reducing the number of parameters to improve the model accuracy.
In this study, electrical conductivity (EC) from Veris iScan sensor and Dualem scanner, yield data, historical yield data from a combine harvester (JohnDeere), and S2 imageries were collected for 10 winter wheat (Triticum aestivum L.) fields in Germany and 10 fields in the UK for the 2020 and 2021 season. The combine harvester data were analysed using ArcGIS software. The topographic wetness index (TWI), a good indicator for soil moisture, was calculated based on the digital elevation model (DEM) using ArcGIS software. Samples for soil organic matter and unmanned aerial vehicle (DJI Mavic 2 Zoom) imageries were also collected for the UK and the German fields respectively. The unmanned aerial vehicle (UAV) was equipped with a compact multispectral sensor (Parrot Sequoia+) and flew 60 m above ground level. The UAV data were analysed in Pix4D software. S2 imageries with 10 m high spatial resolution and 5-day temporal resolution were downloaded from Europe’s Copernicus website. The obtained imageries were analysed using SNAP toolbox.
This study aims to develop a Machine Learning Approach (MLA) based on BN model to predict wheat yield using two novel parameters which are Prior Inherent Potential (PIP) and Inherent Potential (IP). The model has been developed using Netica (Norsys software), categorizing each node within each field from high to low PIP based on the data available for a given field. A high PIP typically leads to a higher IP and a high IP lead in turn to a high yield. Yield predictions are based on the probabilities of 25%, 50%, 75%. This model approach provides promising predictions.
