Classifying raw irregular time series (CRIT) for large area land cover mapping by adapting transformer model

• DOI: 10.1016/j.srs.2024.100123
• Publication Date: 2024-02-09T06:45:37Z
• Authors: Hankui K. Zhang; Dong Luo; Zhongbin Li
• Keywords: Land Cover

For Landsat land cover classification, the time series observations are typically irregular in number of a period (e.g., year) and acquisition dates due to cloud variations over large areas plan long periods. Compositing or temporal percentile calculation usually used transform regular variables so that machine deep learning classifiers can be applied. Recognizing composite calculations have information loss, this study presents method directly Classifying Raw Irregular Time (CRIT) ('raw' means good-quality surface reflectance without any derivation) by adapting Transformer. CRIT uses day year as classification input align also takes satellite platform (Landsat 5, 7 8) address inter-sensor differences. The was demonstrated classifying analysis ready data (ARD) acquired across one for three years (1985, 2006 2018) Conterminous United States (CONUS) with both spatial availability. 20,047 training 4949 evaluation 30-m pixel were where each annotated seven classes year. compared 16-day percentiles 1D convolution neural network (CNN) method. Results showed trained samples had 1.4–1.5% higher overall accuracies less computation than composites 2.3–2.4% percentiles. advantages pronounced developed (0.05 F1-score) cropland (0.02 mixed boundary pixels. This reasonable only on average 7.02, 16.49 15.78 good quality years, respectively, contrast 7.89, 27.72, 26.60 raw series. CNN not simply filling positions no zeros while masking mechanism rule out their contribution. take coordinates DEM which further increased 1.1–2.6% achieved 84.33%, 87.54% 87.01% 1985, 2018 classifications, respectively. maps shown consistent USGS Land Change Monitoring, Assessment, Projection (LCMAP) maps. codes, made publicly available.