Determination of crop types based on remote sensing data using artificial intelligence methods

One of the important subtasks for estimating and forecasting yields is crop mapping. In recent years, remote sensing data has been actively used to solve it, which allows us to quickly obtain information about the state of fields, as well as artificial intelligence methods. The purpose of this work was to investigate the possibilities of using neural network methods to determine crops of agricultural plants based on remote sensing data. Two different data sets are taken as a basis: an open dataset of PASTIS satellite images, as well as a mosaic of aerial photographs of the Agrophysical Research Institute obtained in the fields of the Leningrad region using the Geoscan-401 unmanned system. Five segmentation models (U-Net, U-Net 3+, DeepLabV3, FCN, Swin Transformer) were used for training and their performance was evaluated on a set of satellite image data. The results of the experiment showed that the accuracy of the U-Net 3+ and U-Net models significantly exceeds other models. At the same time, the transfer of models trained on low-resolution satellite images to high-resolution aerial photographs for further training has effectively improved the performance of models.