الفهرس Only 14 pages are availabe for public view
 |  | from | 119 |  |  |
| | | from | 119 | | |
Abstract
Classifying stress and healthy crops is a major concern in agriculture. Traditional techniques are destructive, time consuming and cover small areas. Remote sensing tackled the previous problems by capturing unique signatures for each type of stress. By calculating the similarity value between an unknown stress signature and a reference stress signature, the stress type can be identified. In case the similarity value exceeded certain threshold, then there is a match.
Similarity measures are manually tuned. To automate the similarity process, Learnable Similarity Measures have been proposed. These measures are designed for Hyperspectral sensors. Such sensors acquire super fine information across the electromagnetic spectrum. As a result, they capture a unique signature for each earth object. The proposed approach classifies hyperspectral images by constructing effective feature vectors derived from hyperspectral signatures.
The proposed approach includes two versions. In the first version, the feature vector is obtained by calculating the cosine similarities between: 1) the second order derivatives of the target spectral signature and 2) the class means. In the second version, the feature vector is obtained by calculating nine popular similarity measures between: 1) the target spectral signature and 2) the class means. SVM is then applied to the classification task. The role of SVM is learning similar and dissimilar spectral patterns to act as an adaptive threshold. For both versions, variants using decomposition into five sub-band regions are also proposed.
In the experiments with a benchmark dataset, the proposed approach, especially with nine similarity measures, outperforms other existing methods. This is because it combined various discriminatory features such as orthogonal projections information, correlation coefficients, and many others. The proposed approach results are statistically significant.