Source-linked AI summary
Supervised Classification Performance of Multispectral Images
K. Perumal, R. Bhaskaran
TL;DR
Remote sensing classification requires effective methods for increasingly large and complex imagery, where traditional algorithms have shown weaknesses. The paper compares supervised and unsupervised classifiers on multispectral data and reports that Mahalanobis classification performed best, with 99.7884% overall accuracy and a 0.9716 kappa coefficient in the reported assessment.
Problem
Increasing spatiotemporal dimensions of remote sensing data have exposed weaknesses in traditional classification algorithms, motivating further research into efficient classifiers.
Method
The study compares different supervised and unsupervised classification methods on multispectral remote-sensing imagery.
Results
Mahalanobis classification was the most accurate, achieving 99.7884% overall accuracy and a 0.9716 kappa coefficient in the reported assessment.
Takeaways & Limitations
The findings highlight the importance of considering the relationship between the dataset and classifier when performing successful image classification.
Takeaways & Limitations
Further studies are required to improve classifier use and increase applicability, including algorithms for more classes and land-cover types.
Abstract
from arXiv · showhide
Nowadays government and private agencies use remote sensing imagery for a wide range of applications from military applications to farm development. The images may be a panchromatic, multispectral, hyperspectral or even ultraspectral of terra bytes. Remote sensing image classification is one amongst the most significant application worlds for remote sensing. A few number of image classification algorithms have proved good precision in classifying remote sensing data. But, of late, due to the increasing spatiotemporal dimensions of the remote sensing data, traditional classification algorithms have exposed weaknesses necessitating further research in the field of remote sensing image classification. So an efficient classifier is needed to classify the remote sensing images to extract information. We are experimenting with both supervised and unsupervised classification. Here we compare the different classification methods and their performances. It is found that Mahalanobis classifier performed the best in our classification.
1 INRODUCTION
Remote sensing classification produces thematic maps from imagery for environmental, military, civil, and resource-management applications. Because classifier performance depends on the data and method selected, comparative evaluation and experimentation are important.
- Remote sensing imagery supports mapping and classification of land-cover features such as vegetation, soil, water, and forests.
- Image classification assigns pixels to user-relevant informational classes by comparing them with one another and with pixels of known identity.
- Classification may be the final analytical product or an intermediate step in land degradation, landscape modeling, resource management, and environmental monitoring.
- Vegetation mapping with remotely sensed data substitutes for expensive, time-intensive field surveys across large spatial areas.
- Classifier performance depends on the data, making data understanding, comparative studies, experience, and experimentation important for successful classification.
2 SUPERVISED CLASSIFICATION
Supervised classification methods address multispectral data classification using training data to assign image pixels to classes. Classification quality depends strongly on training-site quality and the sequence of training, signature extraction, and image classification.
- Supervised methods have been developed to tackle multispectral data classification, including statistical, artificial-intelligence, and ensemble approaches.
- The quality of supervised classification depends on the quality of the training sites.
- Supervised classification follows three operations: defining training sites, extracting signatures, and classifying the image.
- Selecting more training sites can improve classification results and support accuracy and interpretation of the classified image.
3 MATERIALS AND METHODS
The study evaluates classifiers using multispectral land-cover imagery from the IRS-P6 LISS-3 sensor over Madurai, Tamil Nadu. A field-survey reference image supports accuracy assessment, and random sampling selects training and testing pixels.
- The experiment uses IRS-P6 LISS-3 multispectral imagery covering Madurai, Tamil Nadu, with 2282 x 2507 pixels.
- The sensor data include red, green, blue, near-infrared, and short-wave-infrared bands.
- A referenced image produced from field surveys is used to assess classification accuracy.
- Random sampling selects pixels for training and testing the classifiers.
4 IMPLEMENTATION
The implementation compares several supervised classifiers on multispectral imagery, including rule-based, distance-based, Bayesian, spectral-angle, neural-network, and covariance-based methods. The study uses ENVI 4.2 and evaluates the resulting classified images.
- The study evaluates different classification algorithms using multispectral data in ENVI 4.2.
- 4.1 Parallelepiped Classifier: The parallelepiped classifier compares candidate-pixel values with upper and lower training limits across two image bands.
- The minimum-distance classifier assigns each candidate pixel to the class whose mean vector has the smallest spectral distance.
- 4.3 Maximum Likelihood: Maximum likelihood estimates class means and variances to calculate probabilities while accounting for brightness variability.
- 4.4 Spectral Angle Mapper: Spectral Angle Mapper measures similarity between pixels and reference spectra using an n-dimensional angle.
- The Mahalanobis classifier uses covariance information and assumes normally distributed band histograms.
5 PERFORMANCE
The study compares classifiers using overall accuracy, kappa coefficient, and confusion matrices, finding Mahalanobis classification most accurate for the evaluated area.
- 99.7884% overall accuracy and 0.9716 kappa coefficient were obtained for the Mahalanobis classification.
- Mahalanobis classification was determined to be the most accurate among the evaluated classifiers.The authors attribute this partly to filtering shadows and classifying highly varied clusters.
- The minimum distance classifier had the lowest accuracies in the comparison.
- The Mahalanobis classifier showed the highest accuracy assessment for this particular area.
6 CONCLUSIONS
The paper concludes that Mahalanobis classification outperforms even advanced classifiers while emphasizing the importance of matching classifiers to datasets. It also identifies broader applicability and land-cover coverage as areas for further algorithm development.
- The Mahalanobis classifier outperformed even advanced classifiers in the comparison.
- Successful image classification depends on considering the relationship between the dataset and classifier.
- Further studies are required to improve classifier use and increase the applicability of these methods.
- New algorithms are needed to classify more classes and broader land-cover categories.