Semantic Segmentation of Aerial Images Using Deep Learning

Semantic segmentation is a computer vision function that incorporates classifying each pixel in an image into a precise category or class. Unlike classification functions where the goal is to assign an exclusive label to the whole image, semantic segmentation furnishes a more detailed concern of the image by segmenting it into regions corresponding to distinct objects or structures. Similarly, this paper represents the model which scheme for utilizing the datasets of spatial resolutions also site for teaching a completely convolutional neural specification named the U-Net to finish the segmentation of aerial images. The dataset comprises 72 images captured by MBRSC (Mohammed Bin Rashid Space Centre) satellites, organized into six larger tiles. Each image is remarked with pixel-sharp semantic segmentation across six distinct classes: Vegetation, Water, Land (unpaved area), Road, Building and Unlabeled. Given the variability in image sizes, a crucial preprocessing step involves resizing the images to dimensions divisible by 256 and then extracting manageable patches. This ensures uniformity, enabling seamless processing and analysis. The accompanying mask images, initially provided in RGB format with HEX color codes, undergo conversion to standard RGB values. Subsequently, these RGB values are translated into integer labels, which are further transformed into a one-hot encoded format. This transformation is essential for facilitating the learning procedure of the deep learning model. The core of the project employs a U-Net architecture, a type of convolutional neural network (CNN) prominent for its effectiveness in image segmentation tasks. U-Net's structure, characterized by its symmetric contracting and expansive paths coupled with skip connections, allows for precise localization and classification of each pixel in the image. The network processes the preprocessed patches, generating segmented outputs that categorize each pixel according to the predefined classes. Post-segmentation, the predicted segmented images are converted back to their original RGB color representations. This step ensures that the segmented output aligns with the initial visual characteristics of the input data. The final phase of the project addresses the assembly of these segmented patches back into their respective larger tile structures. This involves merging the predicted tiles, taking care to minimize blending artifacts through smooth blending techniques. This approach ensures a seamless and coherent larger image, preserving the visual and spatial integrity of the original data.