Published Paper Details:

With the improved availability of high-resolution satellite imagery, it is possible to detect detailed structures on our planet's surface. Expert knowledge, supervision, and fieldwork are required for traditional mapping approaches. This study examines the potential of machine learning methods such as artificial neural networks (ANN), support vector machines (SVM), and random forest (RF), as well as different deep-learning convolution neural networks (CNNs) for high-resolution satellite imagery using optical data from the satellite and topographic factors. To do multi-label categorization of Amazon satellite images, this uses a Convolutional Neural Network (CNN) model. It starts with a CNN model that obtains an F-score of 0.84. Then, using three deep CNN architectures that have recently performed well in the ImageNet Challenge, demonstrate that a ResNet-50 model can get a 0.91 F-score. The results reveal that without any additional post-processing steps, state-of-the-art methods outperform by 9.8% on the Intersection over Union metric. The DeepGlobe Building Extraction Challenge asks participants to extract all building polygons from satellite images. They use the segmentation algorithm Mask R-CNN and a method that combines Mask R-CNN with building boundary regularization to construct polygons that are accurate and complete. This approach, however, provides better regularized polygons than Mask R-CNN, which is useful in many applications.