Simulated Hyperspectral Dataset of Forest Canopies for Unmixing Problems

NOTE: THE REPOSITORY IS CURRENTLY EMPTY. THE SIMULATED IMAGES WILL BE ADDED AS SOON AS POSSIBLE.

We use a recently developed forest canopy simulator HyperBlend. It takes into account 1) the shape of the terrainand the spectral reflectance of the soil, 2) the sun angle for realistic shadows, 3) tree geometry and the averageleaf angle, and 4) spectral reflectance and transmittance of the leaves. It can simulate multilayered heterogeneouscanopies complete with understory and soil. Each simulated scene then contains trees of varying height that mayshadow each other and they can be partially obstructed by other trees. For leaf spectra, we use the measured spec-tra of ten different tree species from the LOTUS dataset. We use two simulated soil spectra from GSV simulator:dry sand that is very reflective and wet peat that is very absorbing. Used sun and sky spectra are simulated withSSolar GOA. HyperBlend supports up to ten different tree models per scene. As each tree model can be assignedonly one leaf spectra, the available intra-species variation in leaf spectra is used by picking different measurementsfrom the LOTUS data. Inter-species variation is achieved by populating the scene with spectra of varying species.Geometrical variance is achieved by setting up 5 different scene variations.

Each spectral image will cover a 50 m square (the simulated area is larger so that trees outside the imaged area canstill affect the image by casting shadows, for example). For unmixing purposes, first a 1024 pixel square (spatial res-olution) spectral image is generated. Each successive image will have one fourth of the pixel count of the previous(256, 64, 16, and 4 pixel squares). The pixel labels in the highest resolution image are divided to the low-resolutionimages to provide the fraction of each spectral signal to be used as a ground-truth for unmixing. All in all, fivegeometry variations with two soil variations are generated in five spatial resolutions resulting in 50 simulated spec-tral image cubes. An example of generated data (drone image from top of the canopy, and mean spectra of selectedregions of interest) is shown below.