Combination analysis of multispectral and radar satellite data

Abstract: Remote sensing technologies, such as satellite imagery, have proven to be a powerful tool for land cover classification when combined with machine learning algorithms. Depending on which type of sensor is used for the imagery, different properties of land cover classes may be distinguished. Because of this, a data set containing a combination of data from different sensors could potentially further improve the classification accuracy. To determine if adding data from the radar sensor on the satellite constellation Sentinel-1 to data from the multispectral optical sensor on the satellite constellation Sentinel-2 could improve the accuracy of land cover classification, a tool for combining data from both satellites was developed. The classification accuracy using the combined data was then compared to using non-combined Sentinel-2 data with a neural network and a random forest classifier. We found that the random forest classifier produced a higher accuracy than the neural network for both the combined data and non-combined data. The combined data increased the accuracy further compared to the non-combined data. However, the increase produced by the combined data was small and most likely not worth the extra computational power required to implement Sentinel-1 data to Sentinel-2 data.