DEEP LEARNING BASED PATCH-WISE LAND COVER LAND USE CLASSIFICATION: A NEW SMALL BENCHMARK SENTINEL-2 IMAGE DATASET


Creative Commons License

Alp G., Sertel E.

IGARRS 2022, Kuala-Lumpur, Malaysia, 17 - 22 July 2022, pp.1-4

  • Publication Type: Conference Paper / Full Text
  • City: Kuala-Lumpur
  • Country: Malaysia
  • Page Numbers: pp.1-4
  • Istanbul Technical University Affiliated: Yes

Abstract

In this paper, patch-wise land cover and land use (LCLU) classification was performed using the state-of-art ResNet 50 and Inception-ResNet-V2 architecture trained with Stochas- tic Gradient Descent(SGD) and Nadam optimizers. A new dataset was generated for the classification task using Sentinel- 2 images having different patch sizes. The image patches were labeled using CORINE Land Cover (CLC) 2018 map. The dataset has 1961 image patches and it was divided into 1397 training and 564 testing patches during the experiment. Our dataset contains samples labeled with 7 CLC Level-2 classes. While the best training accuracy of 98.0% was ob- tained by Inception-ResNet-V2 trained with Nadam. The best testing accuracy of 93.0% was achieved with Inception- ResNet-V2 by using SGD optimizer.