Optimized Hyperspectral Vegetation Detection Using Lightweight Cascaded DCNN

Abstract

Hyperspectral Image Classification (HIC) is essential for distinguishing surface objects and monitoring materials in remote sensing. However, the large number of spectral bands increases computational complexity and classification time. This study proposes a HIC system that uses Deep Convolutional Neural Networks (DCNN) with spectral band selection strategies. The Indian Pines and Salinas datasets are used for evaluation, employing PCA, LDA, and ILDA for band selection. Performance is assessed using accuracy, recall, precision, F1-score, training time, and classification time. The first phase utilizes a three layered DCNN with PCA for feature representation, achieving 98.20% accuracy on the Indian Pines dataset with 30 spectral bands and a 25×25-pixel window. The second phase introduces a Lightweight Cascaded DCNN (LC-DCNN) with ILDA, enhancing classification accuracy. LC-DCNN+ILDA achieves 99.51% accuracy on Indian Pines and 99.71% on Salinas, outperforming other methods. ILDA proves more effective in selecting discriminative spectral bands than PCA and LDA. In the future, adding bigger datasets with more extensive objects can improve performance for real-time datasets.

Related Papers

• → Hyperspectral imaging and analysis of mixed algae species in liquid media(2012)6 cited
• Susquehanna Chorale Spring Concert "Roots and Wings"(2017)
• → DETERMINING QUALITY REQUIREMENTS AT THE UNIVERSITIES TO IMPROVE THE QUALITY OF EDUCATION(2018)
• → A novel reduced-order method for analysis of hyperspectral images(2021)
• → ИСПОЛЬЗОВAНИЕ ПОТЕНЦИAЛA СОЦИAЛЬНЫХ ПAРТНЕРОВ В ПОДГОТОВКЕ БУДУЩИХ ПЕДAГОГОВ(2024)