High Spatial Resolution Remote Sensing Image Scene Classification using CNN with Transfer Learning

LETICIA SEIJAS; ROCIO GIRALDEZ; AYRTON BETTI; JORGE MÁRQUEZ

Congreso; IEEE ARGENCON 2020; 2020

IEEE Sección Argentina y UTN Resistencia

Traditional HSR remote sensing imagery understanding is based on recognizing pixel-based or object-based ground elements, but this cannot describe the whole content of the scene images and cannot bridge the ?semantic gap? between the low-level features and the high-level semantics. Deep learning algorithms attempt to learn hierarchical features, corresponding to different levels of abstraction. Current progress in deeplearning models, specifically deep convolutional neural network (CNN) architectures, have improved the state-of-the-art in many fields of study including remote sensing scene classification. The choice of a proper network architecture for making strongand correct assumptions about the nature of the input data is still a big challenge. This work presents an implementation of the Convolutional Neural Network (CNN) AlexNet trained on the well-known datasets UC Merced Land-Use and WHURS by using Transfer Learning for the High Spatial Resolution Image Scene Classification problem. A layer corresponding to the Spatial Pyramid Pooling (SPP) is incorporated in order to consider different image sizes in the input of the network and multi-scale spatial information of the scenes to be classified. Results are comparable to literature, improving some published approaches.