Abstract
In this study, a 1-D Convolutional Neural Network (CNN) architecture was developed, trained and utilized to classify single (summer) and three seasons (spring, summer, fall) of hyperspectral imagery over the San Francisco Bay Area, California for the year 2015. For comparison, the Random Forests (RF) and Support Vector Machine (SVM) classifiers were trained and tested with the same data. In order to support space-based hyperspectral applications, all analyses were performed with simulated Hyperspectral Infrared Imager (HyspIRI) imagery. Three-season data improved classifier overall accuracy by 2.0% (SVM), 1.9% (CNN) to 3.5% (RF) over single-season data. The three-season CNN provided an overall classification accuracy of 89.9%, which was comparable to overall accuracy of 89.5% for SVM. Both three-season CNN and SVM outperformed RF by over 7% overall accuracy. Analysis and visualization of the inner products for the CNN provided insight to distinctive features within the spectral-temporal domain. A method for CNN kernel tuning was presented to assess the importance of learned features. We concluded that CNN is a promising candidate for hyperspectral remote sensing applications because of the high classification accuracy and interpretability of its inner products.
Highlights
Land-cover maps provide information for natural resource and ecosystem service management, conservation planning, urban planning, agricultural monitoring, and the assessment of long-term land change
Overall accuracy for three-season data was 1.9 to 3.5% significantly higher than single-season data for all three classifiers (Z > 2.1, p < 0.05). Given their superior performance, remaining results will focus on three-season Convolutional Neural Network (CNN) and Support Vector Machine (SVM) classifications
We found that the 1-D SVM and CNN classifiers had over 7% overall accuracy for improvement over theclassifiers
Summary
Land-cover maps provide information for natural resource and ecosystem service management, conservation planning, urban planning, agricultural monitoring, and the assessment of long-term land change. The automated classification of land cover from satellite imagery is a challenging task due to spectral mixing, intra-class spectral variability, and low spectral contrast among classes. Hyperspectral, or imaging spectroscopy, data consist of hundreds of spectral bands, and capture more spectral detail and variability relative to conventional multispectral sensors used for mapping land cover. Terrestrial hyperspectral applications have shown success in mapping composition, physiology, and biochemistry of vegetation, ecosystem disturbance, and built-up environments [1]. The analysis of hyperspectral data presents issues in classification, due to large data volumes, and increased spectral variability as recorded by hundreds of correlated bands [2].
Sign-in/Register to view highlights & summary Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a callDisclaimer: All third-party content on this website/platform is and will remain the property of their respective owners and is provided on "as is" basis without any warranties, express or implied. Use of third-party content does not indicate any affiliation, sponsorship with or endorsement by them. Any references to third-party content is to identify the corresponding services and shall be considered fair use under The CopyrightLaw.
