Hyperspectral imagery contains many diferent frequency bands for each pixel, enabling advanced detection algorithms. The so called pushbroom sensor technology is the most widespread realization of hyperspectral camera. Such cameras capture a single row of pixels, but each pixel can have hundreds of associated frequency bands . Assembling a full image from these pixel rows typically requires having a precise knowledge of the movement of the camera in space while it scans the desired area. In this research we present a procedure to build hyperspectral images from line sensor data without camera position information. The approach relies on an accompanying conventional camera and exploits the homographies between images of planar scenes for image mosaic construction. The hyperspectral camera is geometrically calibrated by using a special target and a variation of the Direct Linear Transform algorithm. This enables mapping the hyperspectral lines to each of the images in the mosaic, and a hyperspectral image can be built using the same homographies computed for the mosaic. Ideally, this research can enable small UAVs to perform enhanced detection and navigation previously reserved for high altitude aircraft.

Mapping the hyperspectral line to a conventional camera

Without knowledge of the plane depth relative to the cameras, it is not generally possible to predict which line of the conventional camera contains each of the world points that the line camera captures. This arises from epipolar geometry. Consider a pair of cameras facing the same direction, such as a stereo array. Given the projection of a world point on the hyperspectral camera image plane, it is possible to predict the line on the conventional camera image plane onto which it is also projected. This line to line map generally depends on the depth of the world points in the camera frames. This dependency is eliminated only if the hyperspectral camera line is forced to coincide with an epipolar line.

Assemblying the hyperspectral data cube

Using the line-to-line correspondence between cameras, as well as the scaling factors and the displacements, the positions of the hyperspectral camera points projected onto the conventional camera can be recovered. Using these new point coordinates for the hyperspectral data and applying the homographies from the previous step, a single hyperspectral image with several frequency band layers is recovered. To achieve this, each of the projected point positions is recalculated to a place in the final hyperspectral image.

Experimental results

We used a piloted, radio-controlled airplane to perform the data collection.  We employed a Sony XCGV60E monochrome conventional camera and a custom-made hyperspectral camera manufactured
by Surface Optics, based on the AVT Prosilica GC 655. Both cameras have 640 pixels of horizontal resolution. The conventional camera has 480 pixels of vertical resolution.

Below is image mosaic overlaid by their corresponding hyperspectral images. Normalized Difference Vegetation Index (NDVI) indicates healthy plants in red, nonvegetable in blue, and ranges between. Pseudocolor was generated by pulling red, green and blue bands from the spectrum.

Research products