Locality Preserving Endmember Extraction for Estimating Green Algae Area

We present a novel endmember extraction method for estimating green algae areas in Geostationary Ocean Color Imager (GOCI) multispectral images. We commence by conducting spectral dimension reduction via the locality preserving projection (LPP) that captures image spatial local characteristics. The LPP enables each projected pixel to encode local information around the original pixel. We then exploit minimum noise fraction (MNF) for reducing the noise effects from the projected pixels. We apply the N-FINDR strategy to the projected pixels for endmember extraction. We compute the abundance of each pixel with respect to the green algae endmember. The sum of all pixel abundances results in the green algae area. In contrast to most existing endmember extraction methods which address pixels separately without considering the local pixel neighboring relationships, our method encourages each endmember to encode the local neighboring information. As salient green algae tends to appear across adjacency multispectral pixels rather than in individual separate ones, our locality preserving endmember extraction method is able to characterize the green algae endmember in a more comprehensive manner and thus results in accurate green algae area estimation. Experimental evaluation validates the effectiveness of our method.