Pap_01_Semantic-Segmentation-of-High-Resolution-Satellite-images.pdf

High-dimensional geospatial data visualization has gained much importance in recent decades. But to analyze it, traditional technologies used in machine learning are not convincing enough, and thus to switch to a subdomain of machine learning called deep learning that has gained popularity because of its accuracy and high dimensional data analysis power. Its convergence with geospatial data analytics shall prove to be a boon to the researchers working in the domain of geospatial data. Though Geospatial information is mostly used in the global mapping process of satellite images. The heterogeneity of the data makes it infeasible for global scale mapping. Therefore, to handle this problem is to partition the entire world into several regions. Semantic segmentation is one such technique and is widely used for information extraction from satellite images. The technique essentially refers to segmenting the input image pixel into multiple semantic regions, that is, to assign a semantic pixel category to each pixel in the image. In this context, we propose a semantic segmentation method that utilizes the spatial information of the high-resolution remote sensing data. The aim is to leverage the openly available data to automatically generate a larger training dataset with more variability and can be used to build more accurate deep learning models. The proposed automatic extraction can capture context information and its symmetric expanding path enables precise localization. The most characteristic property is the upsampling part that has feature channels that allow propagation of context information to higher resolution layers and makes the expansive path roughly symmetric to the contracting path yielding a U-shaped architecture. Mean IOU (mIOU) is used as the performance matrix and results yield 0.79. Since the model is trained on a small training dataset, that makes the deep learning model prone to overfitting. Training on such a small set of images makes this a challenging task. Validation dataset metrics obtained after training will signify the model’s general adaptability on other datasets of other segmentation tasks