| dc.description.abstracteng | Summary
The efficient monitoring of forests is essential in the fields of forest resource assessment, management, and scientific decision-making. Integrated remote sensing imagery and field observations are widely used and are considered highly efficient in the field of forest monitoring (such as the monitoring of forest stocks, biomass, forest carbon, and forest type maps) on a local, regional and even global scale.
In the field of forest monitoring by means of remote sensing, optical satellite imagery plays a critical role. However, for optical satellite image processing techniques applied to high-precision forest monitoring to be truly effective, a number of challenges still exist, such as the variability of the atmosphere, topography, illumination conditions and scale issues. To mitigate or eliminate the impact of these challenges, under the technical framework of the Lin4Carbon project, several refined image processing techniques have been proposed and have also been proven to be quite advanced techniques. This thesis consists of five manuscripts. Each manuscript is inspired by problems encountered in the implementation of the Lin4Carbon project. It is worth noting that in order to demonstrate the universality and advancement of these techniques, the locations and data are not entirely limited to the Lin4carbon project.
Manuscript I proposes a simple and straightforward, moving window-based, rotation-correction topographic normalization model as the means to achieve improved forest mapping. An underlying assumption is that the same forest type has stable spectral characteristics, which are essential if forest monitoring is to be accurate. Our proposed model yields less intra-class heterogeneity and without the “overcorrection” phenomenon in vision when compared to uncorrected data or global correction methods.
Manuscript II further discusses the shortcomings of various topographic normalization methods based on global parameter estimations. In general, the band-specific methods lead to overcorrection, while land-cover-specific methods have difficulties in obtaining prior data on land types, and a variety of newly-advanced topographic normalization methods have difficulty achieving satisfactory results, especially when these methods are applied to sites with large areas and significant differences in landscape. Thus, this manuscript aims to establish a simple, generalized, standard, topographic normalization model for large-area optical satellite images. We assume that the relationship between the spectrum signature and the illumination conditions is specific to the site. Also, the empirical parameters of each pixel can be computed directly from a given window size. Compared with Manuscript I, this manuscript was given the same conclusions in the case of the use of new Landsat 8 images and new illumination conditions. In addition, we further tested our proposed model’s performance on the classification of forest types. We conclude that the proposed method achieves higher classification accuracy and requires fewer training samples.
Manuscript III aims to help understand the mechanism of the dynamic effects of illumination conditions on the normalized difference vegetation index (NDVI), within one year, in mountain forests. The NDVI is widely used to assess forest cover, forest biomass, forest carbon, etc. As such, the quality of an NDVI is directly related to output precision. In steep mountain forests, the issue of the variability of NDVI caused by topography appears not to have attracted much attention. In these regions of high topographic variability, the intra-annual illumination conditions vary significantly from the dynamics of the sun-terrain-sensor geometry. These variances, in turn, affect the spectral response and the derived vegetation indices. The seasonal variations of forest NDVIs in rough terrain are studied in this manuscript. We conducted a statistical analysis of random samplings from May 2013 to October 2014, of all available cloud-free NDVI images of Landsat 8 OLI. We studied how illumination conditions (IL) affect intra-annual NDVI on deciduous and coniferous forests. The findings indicate that IL and NDVI have significant positive linear correlations, and the slope coefficients of linear functions are U-shaped over the course of a year. Meanwhile, we found a positive linear correlation between IL heterogeneity and NDVI variability. Thus, the effects of illumination conditions on NDVI or NDVI-related estimations should be taken into account in both forest monitoring and the quantitative analysis of mountainous areas.
Manuscript IV aims to develop a robust and straightforward haze removal method for optical satellite images. In addition to cloud contamination, multispectral remotely-sensed images are often degraded by haze, which in turn reduces visual interpretability and affects further image analysis. Thus, haze detection and removal techniques are essential to multispectral image preprocessing. We successfully removed the haze from the Landsat 8 OLI data in the project Lin4Carbon area. We also found an improved method to achieve better performance in very high spatial resolution images, which we describe in this manuscript. Unlike the existing haze thickness map-based (HTM-based ) method, the proposed method estimates the HTM from the blue band’s mean vector L2-norm for each pixel, using a given window size. Also, we improved the compensation strategies for both haze and haze-free pixels. The proposed method has been successfully applied to a variety of very high-resolution optical satellite imagery with complex haze coverage in densely built-up areas.
Manuscript V aims to reconstruct a higher spatial resolution multispectral image from a so-called “pan-weighted multispectral reconstruction” approach. To some extent, multispectral data is one of the more critical data sources for various index and index-based forest variable estimation, forest type classification, etc. compared to panchromatic images. Interpolation and pansharpening methods are commonly used to improve the spatial resolution of multispectral data. However, the disadvantages of these methods are also evident, in that traditional interpolation methods result in loss of spatial detail, and pansharpening methods tend to suffer from spectrum loss of fidelity. Thus, we attempt to reconstruct higher spatial resolution multispectral images through a method involving joint panchromatic and multispectral images. The proposed method is characterized by the fact that the reconstructed multispectral image pixels inherit the spatial details of the neighboring pixels of the panchromatic image. Through quantitative analysis and visual comparison, it shows that the proposed method produces better performance than traditional interpolation methods. Manuscript V also demonstrates two potential applications, namely refined pansharpening and NDVI calculation. This method would be an alternative or improved upsampled interpolation method, which may become a widely accepted technique used to refine and preprocess satellite images.
Overall, this thesis attempts to solve several of the technical problems encountered in the implementation of the Lin4Carbon project on optical satellite image processing. The data availability and application potential of optical satellite images are often limited by bottlenecks, including their susceptibility to variations from the atmosphere, terrain and illumination conditions. Once these bottlenecks are overcome, optical remote sensing will be more likely to be widely used in forest monitoring. All of the methods presented in this thesis are concerned with the refinement of image pre-processing techniques, which are in turn expected to have a wide range of applications and are applicable to various optical satellite images. | de |