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

Science Press Release Distribution Services

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

August 4, 2022 Science and Technology 0

The issue of change detection in high-resolution (HR) satellite pictures is addressed in this work. Bayesian active learning disagreement (BALD), an active learning (AL) technique, is used to World view photos of the Greek island of Crete that show urban and suburban regions. The BALD acquisition function may be used while thinking about the classification job because it is based on Bayesian uncertainty. The pool data points that should be selected, according to the BALD principle, are those that are anticipated to maximise the knowledge gathered about the model parameters. In reality, the model shows that the data points that maximise the BALD acquisition function are typically unclear. Importantly, each stochastic forward run over the model would provide the greatest probability assigned to a certain class. In the experiments, results from random sampling (RS) on AL are compared. Investigated are several situations for selecting different numbers of pictures from the training set of a convolutional neural network (CNN). The validation accuracy of the BALD algorithm’s categorization of data as altered or unchanged is superior to that of the RS algorithm, the results show. In fact, the BALD method achieves zero test error compared to the RS algorithm’s test errors of 34.6 percent and 38.5 percent. In actuality, as the quantity of training photos grows, so does the accuracy. In further work, intriguing experiments might be carried out inside the AL acquisition function architecture using estimators from robust statistics. Up till now, no other literature research proving the application of

Author(s) Details:

Lemonia Ragia,
Information Management Systems Institute, Athena Research Center, Artemidos 6, Marousi 151 25, Greece.

Antigoni Panagiotopoulou,
Information Management Systems Institute, Athena Research Center, Artemidos 6, Marousi 151 25, Greece.

Please see the link here: https://stm.bookpi.org/RDST-V10/article/view/7717

Keywords: Bayesian CNN, deep active learning, BALD, dropout, QuickBird, WorldView, change detection

 

Leave a Reply

Your email address will not be published.