Multi-Resolution Deep Learning for Land Use Applications

Increased spatial and temporal resolution of remotely sensed multispectral imagery is crucial for improved monitoring of land surface dynamics in heterogeneous landscapes undergoing rapid change. Given orbital constraints, satellite imaging sensors such as MODIS and Landsat 8 OLI exhibit tradeoffs between frequent/coarse and sparse/fine scenes, and spatiotemporal fusion techniques have been developed to synthesize images with improved spatial and temporal resolutions from such complementary satellite pairs. In contrast, imagery from manned fixed wing aircraft and UAVs can be acquired both frequently and at high resolution over limited areas. Land surface monitoring would greatly benefit from a capability to combine imagery from these disparate platforms, for which inconsistent or irregular revisit times and variabilities in resolution and spectral bands make existing spatiotemporal fusion techniques insufficient to combine them effectively. nbsp;nbsp;This project will exploit these recent machine learning advances to combine imagery from disparate satellite and airborne platforms, using multi-resolution image time series and transferring fine resolution knowledge gained from higher resolution training images to lower-resolution test scenes. We will test the feasibility of the system to provide improved classification of vegetative land cover and estimations of fractional vegetation cover, particularly for agricultural areas that frequently change on a small spatial scale. During Phase I, we will use an unmanned aerial vehicle (UAV) to make weekly multispectral image collects during the growing cycle of several agricultural crops and combine the scenes with Landsat 8 OLI and Sentinel 2 satellite imagery. We will spatially and temporally subsample the high resolution UAV imagery to simulate imagery acquired from a variety of aerial and additional satellite platforms and compare classifier performance for different spatial resolutions and repeat periods.