基于裸土期多时相遥感影像特征及最大似然法的土壤分类

运用单时相遥感数据进行土壤分类及制图,其数据本身易受到其他因素干扰而出现误差,存在一定的局限性,导致制图精度不高。为了提高制图精度,以松嫩平原林甸县为研究区,利用裸土时期多时相Landsat 8遥感影像、DEM数据和全国第二次土壤普查数据,从所有单时相遥感影像中提取出多种分类特征,按照分类特征类型进行压缩处理,得到新的多时相分类特征,将不同分类特征进行组合并分别进行最大似然法分类,得到不同分类特征组合下的土壤类型图,通过不同土壤类型图精度来判断各分类特征对于制图的影响。研究表明,该文所提取的分类特征均可以实现土壤制图,使用压缩处理后得到的多时相遥感数据分类特征完成制图的精度更高,总体精度达到91.0%,研究可为土壤精细制图提供依据。

Soil classification based on maximum likelihood method and features of multi-temporal remote sensing images in bare soil period

Abstract: Remote sensing technology is an efficient method of soil mapping and classification. By using this method, soil classification is mainly based on spectral reflectance characteristics. A single-phase remote sensing image is often used for soil classification but a single phase image only reflects the current situation within the study area. It is easily disturbed by other factors and thus the results are inaccuracy. In addition, it is impossible to reflect the dynamic change of the soils that might be affected by factors such as human and natural factors. In order to improve the accuracy of mapping, we took Songdian County in the Songnen Plain as the research area and investigated the feasibility of using multi-phase images to classify soil. The study area included 5 types of soils such as chernozem, meadow soil, swamp soil, aeolian sand soil and water. A total of 3 Landsat 8 remote sensing images of bare soil were collected, representing 3 phases. Combined with DEM data and the second national soil census data, soil mapping was conducted. Different classification features were extracted by 1) Kauth-Thomas transformation, 2) spectral index of normalized difference vegetation index, normalized differential moisture index and normalized differential water index, 3) topography features of elevation, slope, aspect and curvature, and 4) multiphase features. The maximum likelihood method was used for soil classification. The J-M distance of training samples was calculated to show if the soils were easily differentiated. A total of 90 training samples were used for training and 1000 samples were used for validation. They included 500 samples of chernozem, 220 samples of meadow soils, 196 boggy soils and 43 blown soils. The DEM data was collected at a 30-m resolution. The results showed that the soils became easier to be differentiated with increasing numbers of features. According to the J-M distance analysis under mono-phase and multi-temporal remote sensing data, the identification of chernozem and meadow soil was the most difficult in all soil types. The main reason is that the spectral curve of the surface meadow soil is similar to the spectral curve of its adjacent soil type. With the increase of the classification characteristics, the information in the classification feature dataset was richer, and the J-M spacing was also increased, indicating that the selected classification characteristics and the constructed classification feature dataset are efficient in differentiating different types of soil based on remote sensing imagery. The multi-temporal remote sensing images based on multiphase features could complete soil classification and mapping with an overall accuracy rate of 91.0%, a Kappa coefficient of 0.865. By mono-temporal images based on multiphase features, the overall accuracy of using remote sensing imagery in 2014 was 86.3%, and the Kappa coefficient was 0.794; The overall accuracy of using remote sensing imagery in 2015 was 90.3%, and the Kappa coefficient was 0.855; The overall accuracy of using remote sensing imagery in 2017 was 88.6% and Kappa coefficient was 0.830. Compared to the single features, the multiphase features could greatly improve soil classification accuracy. The study could provide valuable information for soil mapping by remote sensing data.