Abstract
Hyperspectral images (HIs) and segmentation have become a promising solution for different applications such as the production of thematic maps (TMs) of agricultural areas. However, problems such as the Hughes phenomenon and high demand for computational resources are related to the high number of bands of HIs. This study proposes a hybrid approach of Strength Pareto Evolutionary Algorithm 2 (SPEA2) and Random Forest classifier for producing TMs, aiming at band selection and improvement of average recall of segmentation. In experiments, the proposed approach reduced the number of bands on average from 220 to 30 in the Indian Pines image and from 224 to 42 in the Salinas image. The proposed approach was statistically whether identical or better than other approaches regarding the average recall of segmentation. Therefore, the proposed approach is promising as regards band selection and competitive in segmentation being a potential tool for generating TMs.
Recommended Citation
Saqui, Diego; Saito, José Hiroki; Lima, Daniel Caio; Jorge, Lucio A. de C.; and Ataky, Steve, "Approach based on SPEA2-band selection and Random Forest classifier to generate thematic maps from hyperspectral images" (2019). AMCIS 2019 Proceedings. 7.
https://aisel.aisnet.org/amcis2019/ai_semantic_for_intelligent_info_systems/ai_semantic_for_intelligent_info_systems/7
Approach based on SPEA2-band selection and Random Forest classifier to generate thematic maps from hyperspectral images
Hyperspectral images (HIs) and segmentation have become a promising solution for different applications such as the production of thematic maps (TMs) of agricultural areas. However, problems such as the Hughes phenomenon and high demand for computational resources are related to the high number of bands of HIs. This study proposes a hybrid approach of Strength Pareto Evolutionary Algorithm 2 (SPEA2) and Random Forest classifier for producing TMs, aiming at band selection and improvement of average recall of segmentation. In experiments, the proposed approach reduced the number of bands on average from 220 to 30 in the Indian Pines image and from 224 to 42 in the Salinas image. The proposed approach was statistically whether identical or better than other approaches regarding the average recall of segmentation. Therefore, the proposed approach is promising as regards band selection and competitive in segmentation being a potential tool for generating TMs.