Crop mapping through a hybrid machine learning and deep learning method

Abstract

A major function of satellite imagery is crop identification and separation, which is of key importance in acreage estimation, yield estimation, and agricultural management. This study aimed to examine the capability of the object-based method and machine learning and deep learning algorithms in combining optical and radar images to generate crop-type maps. For this purpose, Sentinel-1 and Sentinel-2 satellite images were used for the 2019 crop year in the northwest of Ardabil, Iran. Image segmentation was performed through the multiscale method and various spectral, textural, and radar backscattering characteristics. The main features were identified using Random Forest feature selection and fed to support vector machine, random forest, and convolutional neural network algorithms. The results of these algorithms were combined via the majority voting method. The results showed that the convolutional neural network algorithm performed better than other methods with an overall accuracy of 88.03% and a kappa coefficient of 85.99%. Meanwhile, the support vector machine method showed poor performance compared to the other two methods. According to the findings, combining the results of these algorithms through the majority voting method yielded the best crop identification result with an overall accuracy of 90.34% and a Kappa coefficient of 88.72%.