Deep Learning Artificial Intelligence (AI) for Improving Classification Accuracy for the National Land Cover Database

Keywords: Artificial Intelligence (AI), Deep Learning, Machine Learning (ML), Land Cover
Session Type: Virtual Paper Abstract
Day: Tuesday
Session Start / End Time: 3/1/2022 09:40 AM (Eastern Time (US & Canada)) - 3/1/2022 11:00 AM (Eastern Time (US & Canada))
Room: Virtual 36

Abstract

Land cover change (LCC) has been recognized as one of the most important indicators in the study of ecological and environmental changes. The National Land Cover Database (NLCD) is a critical resource for monitoring those changes that occur on the landscape within the conterminous United States. NLCD has used traditional pixel-based methods, including classification and regression tree (CART), since 2001. However, with pixel-based classifications surrounding pixels are not taken into consideration, which leads to higher spectral heterogeneity. This could result in adjacent pixels being incorrectly labeled as different classes and a noisier classification. Deep-learning-based methods provide an end-to-end solution by using both spectral and spatial information. They can generate more accurate classification than the traditional pixel-based methods, especially for the vegetation classes. Our new methodology is based on a regression approach that uses deep-learning Artificial Intelligence (AI). To test this method, we selected sites with diverse landscapes across the conterminous United States. By comparing the confusion matrix values from AI and CART, we generally found that the accuracy from AI in most NLCD classes was slightly higher than from CART. When visually comparing the results, AI outputs are generally less noisy and the patches more spatially coherent than those from CART classification.