본문 바로가기
  • Home
Basic Search

The Monitoring of Land Cover Changes Using Time-Series Images of Sentinel-2

Jeong, Jong Chul 1

1남서울대학교

Accredited

ABSTRACT

This paper presents a comparison of time-series land cover changes from Sentinel-2 imagery from 2017 to 2018 using machine learning. For the machine learning point sampling, we collected approximately 3% of the total pixels at an identical 250 points per class, referring to a digital land cover map of EGIS. To compare land cover changes, the SVM(Support Vector Machine) model was used. The results of time series analysis shows that areas of wetland and bare soil increased, agricultural and grass decreased, and the areas of built-up, forest, and water did not change significantly over the two years. The monthly f1-score remained between 0.61 and 0.67 and the f1-score in April was the highest. When accuracy verification was performed using the error matrix, water areas showed the highest accuracy. Classes that mainly occur in the context of vegetation activity were often misclassified, and built-up areas were found to be misclassified with classes that see vegetation activity due to seasonal effects. The characteristics of each class were confirmed using variable importance. On average, R, G, B, and NDVI showed high importance values regardless of seasonal conditions, but NIR, SWIR, and Red Edge bands were seasonally affected. Additional studies are expected to improve accuracy by considering the number of samples relative to the class area, the selection of training areas and the selection of indices.

Statistics

  • 116 Viewed
  • 80 Downloaded

Tools

Issue List

Citation status

* References for papers published after 2023 are currently being built.

Basic Search Advanced Search