Hyperspectral Remote Rensing Image (HRSI) classification based on Convolution Neural Network (CNN) has become one of the hot topics in the field of remote sensing. However, the high dimensional information and limited training samples are prone to the Hughes phenomenon for hyperspectral remote sensing images. Meanwhile, high-dimensional information processing also consumes significant time and computing power, or the extracted features may not be representative, resulting in unsatisfactory classification efficiency and accuracy. To solve these problems, an attention mechanism and depthwise separable convolution are introduced to the three-dimensional convolutional neural network (3DCNN). Thus, 3DCNN-AM and 3DCNN-AM-DSC are proposed for HRSI classification. Firstly, three hyperspectral datasets (Indian pines, University of Pavia and University of Houston) are used to analyze the patchsize and dataset allocation ratio (Training set: Validation set: Test Set) in the performance of 3DCNN and 3DCNN-AM. Secondly, in order to improve work efficiency, principal component analysis (PCA) and autoencoder (AE) dimension reduction methods are applied to reduce data dimensionality, and maximize the classification accuracy of the 3DCNN, but it will still take time. Furthermore, the HRSI classification model 3DCNN-AM and 3DCNN-AM-DSC are applied to classify with the three classic HRSI datasets. Lastly, the classification accuracy index and time consumption are evaluated. The results indicate that 3DCNN-AM could improve classification accuracy and reduce computing time with the dimension reduction dataset, and the 3DCNN-AM-DSC model can reduce the training time by a maximum of 91.77% without greatly reducing the classification accuracy. The results of the three classic hyperspectral datasets illustrate that 3DCNN-AM-DSC can improve the classification performance and reduce the time required for model training. It may be a new way to tackle hyperspectral datasets in HRSl classification tasks without dimensionality reduction.
27 July 2022 in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
Synthetic aperture radar (SAR) satellite has been widely applied in real-time flood monitoring as that they are not affected by extreme weather conditions. However, there is no automatic method to quickly and accurately extract flood areas with SAR satellite images. In this article, a UNet combined with the attention mechanism (UNet-CBAM) method has been proposed for extracting flood submerged areas, and both Longgan Lake and Dahuchi in Poyang Lake Basin are selected as the test sites. Based on the polarization characteristics of two Sentinel-1A data of Poyang Lake, both UNet and UNet-CBAM extraction methods are utilized to extract the flood areas, respectively. Compared with traditional SAR image water extraction methods, simulation results demonstrate that UNet can obtain more satisfactory results in recall, precision, and ${F}_1$ -parameter, but it has no capability to guarantee continuity in edges and small bodies of water. Moreover, our proposed UNet-CBAM method can further improve recall, precision, and ${F}_1$ -parameter, respectively. Specifically, when compared with UNet, its recall is increased by 0.8% and 1.2% while ${F}_1$ -parameter is improved by 0.6% and 0.8%, respectively, in the two test sites.
The objective is to estimate forest above-ground biomass (AGB) considering the vertical structures of tropical forest in Mondah test site. Ten DLR F-SAR P-band fully polarized images are utilized to reconstruct the vertical reflectivity through TomoSAR techniques, and then extract multi-feature parameters from vertical profiles of reflectivity, finally these parameters are used for forest AGB estimation with support vector regression (SVR) approach. The results show that the method we proposed for tropical forest AGB estimation is effective and accurate in our test site.
30 May 2022 in The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences
Soil moisture (SM) is one of the important physical properties of soil and plays an important role in agricultural production. Global Navigation Satellite System Reflectometry (GNSS-R) SM retrieval is based on the concept of receiving GPS signals reflected by the ground using a passive receiver. Although many research has been reported showing the capability of the GNSS-R technique on SM application. Due to the diversity of soil types and the complexity of soil composition, the study of soil permittivity is still attracting many interesting. This paper presented the investigations on characteristic and performances for different semi-empirical soil models, to reveal and verify the relationship between dielectric constant and SM under different realistic soil compositions. This study is helpful to improve the accuracy for the GNSS-R SM retrieval and can be used to other SM related studies.
The errors of the theoretical model and the simplified SINC model in forest height inversion are analyzed. Simulation results show that the higher the tree height, the lower the kz value, the error between the theoretical and simplified SINC model is greater in inverting the forest height. When kz is 0.03 and forest height is 40 meters, the difference of the inversion height between theoretical and simplified SINC model can reach 1.5 meters. To address this problem, bisection iterative algorithm is utilized to eliminate or minimize the error between theoretical and simplified SINC forest height inversion model based on X-band airborne interferometric synthetic aperture radar (InSAR) coherence data. Lidar $H_{100}$ CHM data are used to verify the effectiveness of bisection iterative algorithm. Compared with simplified SINC model, the performance of bisection iterative algorithm is better in retrieving the real value in the theoretical SINC model. The results show that the average height error of the bisection iterative algorithm is 0.5 m lower than that of the simplified SINC model in the inversion of forest parameters in the height range of 15–20 m.