Aerial image segmentation in significant part shapes applications of remote sensing like land cover classification, urban planning, and disaster management. However, the complex and variable nature of aerial photography causes great challenges, particularly with relation to high accuracy and computational efficiency. Conventional segmentation methods can find it challenging to generalise over multiple contexts, which leads to less than perfect performance in particular circumstances. We report a novel approach to address these problems combining Autoencoders with the Non-Dominated Sorted Genetic Algorithm-II (NSGA-II), enhanced by Non-Linear Analysis techniques. While still preserving significant information, good feature extraction with the autoencoder helps to reduce the dimensionality of the input data. Subsequently, NSGA-II maximises the accuracy and simultaneously reduces the processing cost to so maximise the segmentation procedure. Non-linear analysis guarantees that the segmentation method fits the non-linear features inherent in aerial images, therefore enhancing the optimisation process. Based on segmentation accuracy and efficiency, our experimental results clearly reveal that the proposed method greatly surpasses existing methods. The approach especially gets an average accuracy of 94.7%, a precision of 93.4%, and a recall of 92.1% over numerous test sets. Moreover, the approach suits for real-time applications since the computational cost is 27% less than state-of- the-art methods.