Cost-sensitive stacked sparse auto-encoder models to detect striped stem borer infestation on rice based on hyperspectral imaging

Striped stem borer (SSB) usually causes serious damage to rice, and timely detection of SSB infestation is crucial in rice production. Hyperspectral imaging technology has been employed in to detect abiotic and biotic stresses on plant. In this study, the cost-sensitive stacked sparse auto-encoder (SSAE) was proposed to detect the early infestation stage on rice combined with visible/near-infrared hyperspectral imaging technology. The Fisher linear discriminate algorithm (LDA) was modified to quantify the distribution of feature representation of each layer. Multiple structures of the cost-sensitive SSAE were compared and the optimal structure was two depths with width of 6-6 and spares constraint of 0.1. The cost-sensitive SSAE acquired highest total accuracy of 93.44% with full input variables and satisfying total accuracy of 90.98% with reduced input variables, which were superior and stable in comparison with other state of-art feature extraction and selection methods. These results indicated that the cost-sensitive SSAE has great potential in detecting early SSB infestation based on hyperspectral imaging technology. (C) 2019 Elsevier B.V. All rights reserved.