JEOS RP ISSN02 | Page 19

14

J. Eur. Opt. Society-Rapid Publ. 21, 31( 2025)

Table

3. The spatial information retention effect of various dimensionality reduction methods.

Fig. 13. Detection results on the Gulfport dataset.

datasets using different detection methods. For the San Diego-I dataset, the image obtained after dimensionality reduction via SuperPCA resulted in singular values when detected by HUD, leading to detection results filled with zeros. In the Jungle dataset, the image processed by Super- PCA contained Not-a-Number( NaN) values, rendering detection impossible. The DGAE method involves an array with a size quadratic to the number of spatial pixels during dimensionality reduction. Due to the large spatial dimensions of the Cement Street dataset, Holly dataset, and Jungle dataset, the memory required to run the DGAE dimensionality reduction code exceeded capacity, preventing the experiments from being conducted. Additionally, an error occurred during the superpixel segmentation step of SuperULDA dimensionality reduction on the Jungle dataset, resulting in no dimensionality reduction output. From a subjective evaluation of the detection results, SuperPCA, SuperULDA, and DGAE methods performed poorly on most datasets, exhibiting a significant number of false alarms and chaotic detection outcomes. PCA and MNF methods showed fewer false alarms but also had lower detection rates. The proposed method and the LRCRP method performed relatively well, accurately detecting targets in some datasets, though they exhibited