Анализ информативности спектральных и текстурных признаков при классификации растительности по гиперспектральным аэроснимкам

摘要

Рассматриваются результаты экспериментов по совместному использованию спектральных и текстурных признаков для классификации растительного покрова на авиационных гиперспектральных изображениях. Проанализирована информативность текстурных признаков, предлагаемых в пакете ENVI, на различных участках спектра в диапазоне 400-1000 нм. Приведены примеры, в которых совместное использование спектральных и текстурных признаков позволяет повысить точность классификации.%The article discusses the results of experiments on the joint use of spectral and texture features to classify vegetation cover on airborne hyperspectral images with a spatial resolution of 1 m per pixel. Images were obtained by the experimental sensor with 290 bands in the range 400-1000 nm. The most informative bands have been selected for each scene. Different types of forest vegetation were mainly represented in the analyzed scenes. The informative content of the eight texture features (co-occurence measures) from the ENVI package was studied in different spectral ranges and for different window sizes. Statistical separability of training samples and quality of classification in the whole image were evaluated for the spectral and texture features. All the experiments showed that the most informative texture features might be obtained in the near-infrared range when the window had size of 7×7 pixels. Despite the good separability of samples for the texture features, the final results of classification showed that the classification of spectral features in most cases gives the best results. Texture features can improve the classification quality only using a small number of bands. But to boost the informativeness it should be dealt with an individual for each scene set of texture features. The article provides two examples to illustrate this statement.