Field Detection from Satellite Images with Deep Learning Methods

Abstract

Today, effective production is interrupted unless land ownership disputes are resolved. The state cannot make the necessary investments due to these disputes not being concluded, and the borders of the fields remain unclear. Artificial intelligence-based methods can be suggested to eliminate disagreements and uncertainty. By using convolutional neural network (CNN) based deep learning networks in which image data are meaningful, areas with primary importance in crop production have been identified in this study. With the CNN networks used by computer vision technology, meaningful information can be extracted from the image. Field detection processes were carried out in this study by using deep learning networks that learn from data. As remote sensing studies gain speed, the number of deep learning studies also increases. For this purpose, satellite images were first collected from the Google Earth website, and then these collected images were used in Faster R-CNN and SSD training, which gained a reputation for accuracy and speed. It is aimed to provide more efficient production and resolve disputes by detecting the fields from satellite images. From two different networks running, SSD outperformed Faster R-CNN in terms of both accuracy and run time. With an f1 score of %97.32, SSD gave Faster R-CNN %3.18 superiority. In the field object results in the test images, the SSD outperformed by detecting 12 more fields. In terms of run times, the SSD performed faster detections with a difference of 285.5ms in the experiments tried in one-third of the test images.