| Summary: | Plant classification is a science that is used to assess the quality of forest resources and has been studied extensively. In this paper, we proposed a novel convolutional neural network known as the Fourier Dense Network (FDN) which is a data-driven method to classify the optical aerial images of plants. To efficiently classify plants, the FDN learns and extracts the features of plants in time and frequency domains from the optical images captured using an unmanned aerial vehicle (UAV). In FDN, we designed a fast Fourier dense block (FF-dense block) that describes the features of the plants by using the magnitude and phase information in the frequency domain. Moreover, we used a transition layer to reduce the dimension of feature maps between two FF-dense blocks in the time domain. The primary contributions of this study are as follows: (1) an FF-dense block that considers frequency information and transfers the information into various layers of a block was designed; and (2) the characteristics between the time and frequency domains were repeatedly extracted and combined to more effectively describe the characteristics of tree species. To evaluate our study, we established a novel dataset comprising the UAV-based optical images of plants-vegetational optical aerial image dataset-for conducting plant classification and information retrieval. The dataset contains more than 21863 images of 12 plants. To the best of our knowledge, this is the largest publicly available dataset of the UAV-based optical images of plants. The experimental results demonstrated that the FDN can achieve state-of-the-art performance in terms of plant classification.
|
|---|
