Ichihara Naoto

Study on the automatic discrimination of land cover change locations using time-series satellite image datasets

Takahashi Kazuyoshi

This study aimed to examine the automatic discrimination of land cover change locations using a time-series observation dataset. We performed land cover change discrimination using a convolutional autoencoder that learned the features of land cover.
First, we performed land cover change discrimination using difference analysis, a conventional method, as a standard for discrimination results. We also conducted preliminary examinations to improve the reproducibility of output images by the neural network dataset and examined the relationship between land cover change and SSIM. These examinations revealed that the reproducibility of images improves by using land cover information from multiple regions in the learning data, and that the SSIM of the output image after the change decreases compared to before the change when a land cover change occurs.
Based on these findings, we performed land cover change discrimination using a convolutional autoencoder with changed input data and learning data, and verified the discrimination performance of the neural network dataset created.
The results showed that the neural network dataset created in this study did not reach the discrimination results of previous studies and conventional methods. However, the neural network dataset that includes the target area in the learning data approached those methods in discrimination results. This suggests the potential applicability of land cover change discrimination by the neural network dataset created in advance. However, in the case of the neural network dataset that does not include the data of the target area in the learning data, it showed a low discrimination result. This is thought to be due to the fact that there was a large difference in SSIM between the output images before and after the land cover change. Furthermore, the method proposed in this study changes the discrimination results depending on the threshold, so it is desirable to improve the method of setting more appropriate thresholds and the learning data to be used.