Anomaly detection of hyperspectral images based on improved isolation forest algorithm.

Hyperspectral images (HIs) include roughly continuous spectral and spatial data pertaining to land cover, enhancing the ability to identify land cover. Therefore, HIs have significant application value in fields such as remote sensing and environmental monitoring. At present, HI anomaly detection has problems such as low computational efficiency, low detection accuracy and poor adaptability. To solve these problems, the research introduces the Godec method based on the Isolation Forest (IForest) algorithm and combines the methods of global and local evaluation for improvement, addressing the issues of local false detecations and false alarms in traditional algorithms. Meanwhile, based on multi-scale spatial constraints combined with Gabor filters and the Entropy Rate Superpixel (ERS) algorithm, spatial features are extracted from multiple scales, and the abnormal scores in the spatial domain and spectral domain are comprehensively calculated to achieve abnormal target detection. The research conducts experiments using two sets of standard hyperspectral datasets, namely San Diego and HYDICE. The results show that the area values under the curve of the research method on the training set and the test set reach 0.973 and 0.967 respectively, the F1 score is 0.937, and the false alarm suppression rate reaches 92.14%. The detection rate and missed detection rate of the research method in the monitoring of mining area collapse are 95.28% and 4.08% respectively. The detection rate and missed detection rate in military camouflage reconnaissance are 93.79% and 3.82%. The method proposed in the research provides a highly reliable and low-false alarm rate detection tool for fields such as mining area collapse monitoring and military camouflage reconnaissance.