Multi-Source Optimization using Deep Learning framework for Downscaling Satellite-Derived Thermal Data
Land Surface Temperature (LST) from MODIS (MOD 11) has a coarse spatial resolution (1km). As the agricultural lands in India are mostly 200-300m, such images are unsuitable for monitoring the conditions of individual fields. Hence drought-related applications suffer due to lack of temperature data. MODIS-LST are received daily and hence contain the potential variability about temperature conditions. To benefit from this high temporal information, the finer thermal resolution data (e.g. from LANDSAT) need to be embedded but they are available at about 16 days frequency. Therefore, it is important to have LST data at finer resolution more frequently. However, downscaling LST data at finer resolution will have a trade-off between accuracy and the level of the lower pixel size. Since existing techniques for disaggregation do not show superiority of the adjustment of the resolution ratio and are linear. But the ground land cover composition inside a coarse spatial resolution is non-linear, and hence deep learning regression model and fusion algorithms using CNN Autoencoder and LSTM utilizing data at multiple resolutions from Sentinel-2, Landsat-TM/ETM+/OLI and MODIS will be investigated in this study for potential enhancing relationship of LST and other variables.
The broad objective is to propose an efficient image data analytics based on a deep learning framework to predict an accurate downscaled LST at high spatial and high temporal resolutions using LANDSAT-derived LST, SENTINEL 2-derived Vegetation Condition (VC), and MODIS-derived LST. Following are the focused objectives:
• To Predict land surface temperature at 100m spatial resolution from 1km MODIS LST data and optimize the error between the actual LST at 1km to the predicted LST through statistical evaluation at various heterogeneous features.
• To Estimate land surface temperature at 20m spatial resolution using a three-faceted interaction and deep learning model (using data from MODIS, LANDSAT, and SENTINEL-2).
• To propose a suitable spatio-temporal fusion algorithm for blending MODIS, LANDSAT, and SENTINEL-2 data to minimize the trade-off of the uncertainty due to differences in spatial and temporal resolution data.