Slums are created as a result of unprecedented urbanization, especially in developing nations. Remote sensing has shown to be a very useful and efficient tool for mapping these slums. Recent advances in deep learning allow the specific morphological features of slums to be detected even in high resolution remote sensing imagery. The scarcity of available data on slums can be one of the major challenges in detecting these settlement structures, as well as the inter-and-intra urban variability of slums, and their possible similarity to other urban built-up structures. Thus, in our study we aim to address these challenges by adapting a few-shot meta-learning technique to our custom deep learning model STnet. Even when using only very few samples, ranging from 1 to 32 image tiles, we could reach high accuracy rates of up to 74%. We could also reduce the number of parameters in our custom STnet by more than half compared to a typically used Resnet12, while achieving the same accuracies. Few-shot meta-learning proves extremely valuable in retrieving information on complex and heterogeneous urban structures such as slum patches, while also reducing the amount of training data needed.
Paper
Full text
Detecting challenging urban environments using a few-shot meta-learning approach
Semantic Scholar · Computer Science · 2023
Abstract
Slums are created as a result of unprecedented urbanization, especially in developing nations. Remote sensing has shown to be a very useful and efficient tool for mapping these slums. Recent advances in deep learning allow the specific morphological features of slums to be detected even in high resolution remote sensing imagery. The scarcity of available data on slums can be one of the major challenges in detecting these settlement structures, as well as the inter-and-intra urban variability of slums, and their possible similarity to other urban built-up structures. Thus, in our study we aim to address these challenges by adapting a few-shot meta-learning technique to our custom deep learning model STnet. Even when using only very few samples, ranging from 1 to 32 image tiles, we could reach high accuracy rates of up to 74%. We could also reduce the number of parameters in our custom STnet by more than half compared to a typically used Resnet12, while achieving the same accuracies. Few-shot meta-learning proves extremely valuable in retrieving information on complex and heterogeneous urban structures such as slum patches, while also reducing the amount of training data needed.