Active Learning in Deep Bayesian Framework for Detecting Changes in Urban/Suburban Satellite Image Targets of High-resolution

Abstract

This work addresses the problem of change detection in high-resolution (HR) satellite images. The active learning (AL) algorithm called Bayesian active learning disagreement (BALD) is applied on World view images that are depicting urban and suburban areas in the island of Crete, Greece. The BALD acquisition function is based on Bayesian uncertainty and can be utilized considering the task of classification. As the BALD principle defines, the pool data points which should be chosen are those being expected to maximize the information gained about the model parameters. In fact, the model demonstrates on average uncertainty concerning the data points that maximize the BALD acquisition function. Importantly, the highest probability appointed to a certain class would be given by each stochastic forward pass across the model. The experiments present comparisons with results from random sampling (RS) on AL. Various scenarios for picking varying numbers of images in a convolutional neural network's (CNN) training set are investigated. According to the results, the validation accuracy of classification as changed or unchanged of the BALD algorithm is superior to that of the RS algorithm. Indeed, the BALD algorithm achieves zero test error against the test errors 34.6% and 38.5% of the RS algorithm. As a matter of fact, as the amount of training images increases, the accuracy increases as well. In future work, estimators from robust statistics could be utilized inside the AL acquisition function framework to perform interesting experiments. No other literature study has been presented up to now demonstrating the usage of deep AL on WorldView images for change detection.