Early Identification of Potato Fields using Data Fusion and Artificial Neural Network

We propose an approach for the preseason identi-fication of potato fields using a fused dataset that consists of a multispectral image from Sentinel-2 and a Synthetic Aperture Radar (SAR) image from Sentinel-1 satellite. Ground truth data is provided by the National Institute of Potato and Sugar Beet, Brasov, Romania. Both images were acquired on the 23<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">rd</sup> of May 2023, when the characteristic ridges were prepared during the potato seeding. We processed and visualized the temporal data using all the available SAR Ground Range Detected images in 2023 to observe the changes. Further, we aligned the pixels from multispectral and SAR images and created vectors that represent the fused data points where each vector has five values representing features in the dataset. The final dataset has 5233 vectors which are further used to train and test an artificial neural network (ANN) with two hidden layers and it performs binary classification of potato and non-potato fields. As a result, a training accuracy of 94.41% and a test accuracy of 93.03% were achieved. We compared this with results from the crop growth season (96.88% accuracy), noting the increased contribution of the Normalized Difference Vegetation Index (NDVI) during this period compared to bare soil classification.