In order to join virtual sessions, you must be registered and logged-in(Were you registered for the in-person meeting in Denver? if yes, just log in.) 
Note: All session times are in Mountain Daylight Time.

Deep learning-based time series of land use and land cover mapping in South and Southeast Asia

Authors: Yaqian He*, Dartmouth College, Justin S Mankin, Department of Geography, Dartmouth College, Hanover, NH, USA; Department of Earth Sciences, Dartmouth College, Hanover, NH, USA; Lamont-Doherty Earth Observatory of Columbia University, New York, NY, USA, Jonathan Chipman, Department of Geography, Dartmouth College, Hanover, NH, USA.
Topics: Remote Sensing, Land Use and Land Cover Change, Asia
Keywords: land use and land cover change; Convolutional Neural Networks; South and Southeast Asia; NDVI
Session Type: Paper
Presentation File: No File Uploaded


South and Southeast Asia, home to some 33% of the world’s population, have experienced rapid land use and land cover change (LULCC), such as urbanization and deforestation. A spatially-completed and temporally-continuous land use and land cover dataset of South and Southeast Asia is crucial for understanding its LULCC drivers, environmental and climate consequences, and future conservation measures. Remote sensing with repeated data coverage offers an indispensable way to monitor LULCC. Here we evaluate the use of temporal Convolutional Neural Networks (TempCNNs) to document nearly 35 years (1982-2015) of land use and land cover in South and Southeast Asia. We use NOAA AVHRR GIMMS Normalized Difference Vegetation Index third generation (NDVI3g) dataset, with reference data from MODIS) land cover type product (MCD12Q1) and the European Space Agency (ESA) Climate Change Initiative (CCI) land cover product. Our maps demonstrate that Southeast Asia has experienced far more dramatic LULCC compared to South Asia. Cropland expansion in Burma and Thailand has been striking over last three decades, in contrast to cropland and forest decreases in southern China and Indonesia, respectively. These annual maps would be a valuable source for climate and environmental studies in this region. Methodologically, we explore the sensitivity of our classification to our choice of training and reference data, finding that the raw NDVI time series generates more accurate maps than the double-logistic smoothed NDVI time series by some 5 percentage points, and that the MODIS reference data better serves its purpose (over CCI) due to its broader classification system.

To access contact information login