Deep Learning Based Land Cover and Crop Type Classification: A Comparative Study

Remote sensing data is available free of cost with an ever-increase in the number of satellites. This satellite imagery can be used as raw input from which cultivated/non-cultivated and crop fields can be mapped. Previous trends included the use of traditional ML techniques and standard CNN, RNN for such mappings. In this paper, we investigate the segmentation models for the task of Landcover and Crop type Classification. We investigate the UNet, SegNet, and DeepLabv3+ in the data-rich states of Nebraska, Mid-West, United States. We acquire dataset from Cropland data Layer provided by USDA National Agricultural Statistics Service. Our Experimental results show that cultivated and non-cultivated landcover is classified with an accuracy of 90% and crop types are classified around 70% ensuring the models trained on one geographical area can be used for accurate classification in other geographical areas, which makes it more reliable for real-time application in agricultural business. [GitHub]

Automated summary

This study investigates segmentation models for landcover and crop type classification tasks, with experimental results showing good performance of the models in accurate classification across different geographical areas.