Optimal scale of crop classification using unmanned aerial vehicle remote sensing imagery based on wavelet packet transform

Abstract: For the high-resolution remote sensing imagery, the space scale has effect on the classification accuracy and efficiency. The UAV image can achieve very high spatial resolution, which has promoted the development of object-oriented classification, at the same time, it also takes some influence on high-resolution remote sensing imagery classification. Therefore, it's necessary to select optimal scale for classification. In this paper, to resolve high-resolution remote sensing imagery classification problem, we used UAV aerial crop images in Zhuozhou, Hebei province, as a data source, and applied wavelet packet transform to the multi-scale quantitative analysis of classification characteristics that belong to high-resolution remote sensing imagery. Wavelet packet decomposition was applied to seven of the most popular crops images, and then considering multi-factors, the most optimal wavelet packet decomposition tree were selected. Then, we selected the high frequency leaf node of each level according to the optimal wavelet packet decomposition tree and built the texture characteristic vector by four kinds of textural information including mean, variance, energy and energy difference which were computed statistically by the wavelet packet coefficient of selected node. Spectral information was obtained from the low frequency part. The classification characteristic vector was built by integrating the texture characteristic vector and the spectral information. To analyze the separability of the crop sample in different levels of wavelet packet decomposition tree, we needed to calculate the J-M distance of the classification characteristic vector between samples in different levels by matching the wavelet packet levels and the resolution, and then acquiring the optimal spatial scales for object-oriented classification. To verify the result, we conducted object-oriented classification experiment based on wavelet packet transform on imagery of different resolution, and chose accuracy of object-oriented classification and time-consuming of division as evaluation criterion to evaluate the result. The original images were decomposed to five levels, from which wavelet packet transform method was used. The texture information and spectral information which can extract from the optimal wavelet packet decomposition tree were used to build classification characteristic diagram. Then we acquired resampling images of different resolutions which matched with the wavelet packet decomposition levels. Classification characteristic diagram as the thematic layers was used to classify the imagery. Finally we employed overall accuracy, Kappa and time-consuming to assess the suitable scale. The results showed that, 1) In the third and fourth levels of wavelet packet decomposition tree (the spatial resolution was 0.32 -0.64 m), the J-M distance of different samples become maximum which meant the strongest separability; 2) The accuracy of object-oriented classification based on wavelet packet transform was the highest in overall accuracy (0.90 and 0.89) when the spatial resolution was 0.32-0.64 m, and also saved a lot of time than the higher resolution (0.16 m). We concluded that it was suitable for crop object-oriented classification in the third and fourth levels of wavelet packet decomposition (the spatial resolution is 0.32-0.64 m). The method used in this paper for selecting optimal spatial scale for crop classification in high-resolution remote sensing imagery base on wavelet packet transform can accurately select the spatial scale with optimal classification accuracy and the highest classification efficiency. To some extent, the classification accuracy was improved by the classification characteristic which extracted via the method of wavelet packet transform. This proposed method may help with the fine recognition of crops using high-resolution remote sensing images.