Spaceflight and planetary exploration place severe constraints on the available bandwidth for downlinking large hyperspectral images. Communications with spacecraft often occur intermittently, so mission-relevant hyperspectral data must wait for analysis on the ground before it can inform spacecraft activity planning.
A future generation of hyperspectral imagers, such as HyspIRI, will return unprecedented volumes of image data; communication bandwidth constraints and latency will be a key constraint on the total data yield. A similar issue confronts planetary exploration missions that must communicate with Earth over the Deep Space Network. Endmember detection can improve mission science return in both cases. It permits more sophisticated analyses, such as novelty detection, change detection against historical catalogs, scene summary, and data reduction prior to onboard classification to find specific targets of interest. Additionally, endmember analysis can facilitate data summary for downlink. The technology described is suitable for use with onboard hyperspectral instruments and/or their host spacecraft. An efficient superpixel-based endmember detection algorithm keeps to the limited computational constraints of spacecraft flight processors. EO-1 (Earth Observing One) flight experiments demonstrate that the procedure enables significant improvements in downlink efficiency.
While the implementation of both Felzenszwalb graph segmentation and SMACC (Sequential Maximum Angle Convex Cone) endmember detection algorithms is straightforward, additional optimizations were necessary to fit within the limited memory and processing constraints of spacecraft compute environments. In particular, Felzenszwalb segmentation leverages a disjoint-set graph representation with path compression whenever a node is accessed. Path compression amortizes the cost of parent node searches and makes common segment joins operations highly efficient. With this optimization, the dominant computation and storage cost for segmentation is in the initial graph construction, effectively trading memory size for runtime speed. Graph construction involves computing pair-wise distances between spatially adjacent pixels, and several common distance metrics were implemented for use onboard. For each one, nonessential floating-point operations were eliminated. For instance, when computing Euclidean distance, the final square root computation is omitted; for spectral angle distance, only the dot product projection is computed — the inverse cosine is dropped. Such optimizations are justified, since the relative distance between pixels does not change. The computational cost of SMACC endmember detection is dominated by dot product projections. However, since superpixel segmentation has already reduced thousands of individual multispectral pixels to hundreds, by keeping the number of endmembers identified in the tens, the total number of dot product projection operations is limited to a few thousand.