Classification of very high-resolution remote sensing image ground objects using deep learning

In the last decade, advances in remote sensing technology have increased the availability of high-resolution remote sensing imagery that is increasingly reliable, especially in land cover mapping. The complexity of information generated from these high-resolution remote sensing images poses new challenges in land cover classification. This paper proposes a deep learning-based remote sensing image classification approach using a modified Convolutional Neural Network (CNN) architecture called MACNN. The architecture modification is done because manual CNN on remote sensing images could be more efficient, as the architecture was initially proposed for natural image processing. We improved the convolution layer to match the characteristics of remote sensing imagery and added a dropout layer to the convolution block and hidden layer to reduce overfitting. MACNN showed better results than other comparison methods regarding mIoU and mPA, which were 67.23% and 80.08%, respectively. MACNN showed superiority in 8 out of 15 classes compared to other comparison methods for classification results per class. The superior classes include industrial land, rural residential, paddy field, garden plot, shrub land, natural grassland, and lake classes with IoU values of 76.25%, 74.58%, 84.66%, 64.2%, 86.4%, 73.65%, 76.76%, and 93.27%, respectively. Based on several evaluation metrics that have been conducted, MACNN has outstanding accuracy on most metrics and classes compared to existing methods.