Accuracy evaluation of convolutional neural network classification algorithms for building identification in rural and urban areas from very-high-resolution satellite imagery in Jambi, Indonesia

Accurate land cover data are essential to a reliable decision-making process; therefore, researchers have turned to novel land cover classification algorithms employing machine learning on high-resolution satellite imagery to improve classification accuracy. The experiment presented in this paper aims to assess the accuracy performance of three patch-based, convolutional neural network architectures (LeNet, VGGNet, and XCeption) in classifying building footprints in rural and urban areas from satellite imagery data, with conventional, pixel-based classification algorithms as a benchmark. The experiment concluded that the CNN classification algorithms consistently outperformed pixel-based algorithms in the accuracy of the resulting building-footprint classification raster. It was also demonstrated that larger image patch size does not always improve classification accuracy in all CNN architectures. This study also revealed that the XCeption architecture performed best among the three CNN architectures assessed, with a 72-pixel patch size having the best accuracy.