Urban Change Monitoring with Neural Networks and Deep-Temporal Remote Sensing Data

Abstract

Remote sensing driven urban change detection and monitoring has been studied for over five decades. It is applied widely for understanding socio-economic impacts, identifying new settlements, or analyzing trends of urban sprawl. Many methods have evolved to detect such man-made changes and, with the recent advent of Deep Neural Networks (DNNs), have led to a rapid growth of approaches in the past years. The downside of these methods is their dependency on expensive high quality observations that require manual selection and pre-processing. Also adaptability or transfer to different regions and environments imposes non-trivial requirements on the data and methods. These aspects induce limitations in their practicability, scalability, and temporal resolution. In the present work, two supervised methods were introduced to enable urban change monitoring with a high temporal resolution and in-expensive large-scale low- to medium-resolution multi-modal level 1 data. The use of level 1 Synthetic Aperture Radar (SAR) and optical multispectral data required a resiliency of the methods to partial and irregular observations. This was addressed by the use of windowed time series with large but varying amount of observations, which is introduced as deep-temporal remote sensing data. The methods were separated into two stages. First, a DNN was pre-trained with a fully automatic process. It utilized synthetic but noisy labels, derived from the level 1 data, to monitor urban changes in each window. Second, transfer learning was applied to the pre-trained network to customize it for a target location. This was accomplished by providing a smaller set of labels, manually derived from few but publicly available Very High Resolution (VHR) observations. This process only required little manual efforts for the transfer stage through a novel window aggregation method. This simplified labeling still provided convergence during training and improved the overall monitoring performance.Remote sensing driven urban change detection and monitoring has been studied for over five decades. It is applied widely for understanding socio-economic impacts, identifying new settlements, or analyzing trends of urban sprawl. Many methods have evolved to detect such man-made changes and, with the recent advent of Deep Neural Networks (DNNs), have led to a rapid growth of approaches in the past years. The downside of these methods is their dependency on expensive high quality observations that require manual selection and pre-processing. Also adaptability or transfer to different regions and environments imposes non-trivial requirements on the data and methods. These aspects induce limitations in their practicability, scalability, and temporal resolution. In the present work, two supervised methods were introduced to enable urban change monitoring with a high temporal resolution and in-expensive large-scale low- to medium-resolution multi-modal level 1 data. The use of level 1 Synthetic Aperture Radar (SAR) and optical multispectral data required a resiliency of the methods to partial and irregular observations. This was addressed by the use of windowed time series with large but varying amount of observations, which is introduced as deep-temporal remote sensing data. The methods were separated into two stages. First, a DNN was pre-trained with a fully automatic process. It utilized synthetic but noisy labels, derived from the level 1 data, to monitor urban changes in each window. Second, transfer learning was applied to the pre-trained network to customize it for a target location. This was accomplished by providing a smaller set of labels, manually derived from few but publicly available Very High Resolution (VHR) observations. This process only required little manual efforts for the transfer stage through a novel window aggregation method. This simplified labeling still provided convergence during training and improved the overall monitoring performance.