基于RS数据与RF算法的陕西省土壤有机质预测研究

发布时间：2018-04-24 21:16

本文选题：多分辨率遥感数据 + 随机森林算法　；参考：《西北农林科技大学》2016年硕士论文

【摘要】：土壤有机质是土壤最具代表性的属性之一,是评价土壤肥力和土壤质量的一项有效指标。调查采样是了解区域土壤有机质状况的主要手段,但单纯依靠实测样点的土壤有机质预测即费时费力又存在着局部特殊地形地貌区偏差较大的问题。随着3S技术的发展,借助辅助环境因子的土壤属性预测在数字土壤制图中日益受到重视,遥感数据作为反映土壤组成结构及植被生长状况的数据源逐渐被学者们应用到土壤属性的预测之中,并发挥独特优势。本文利用随机森林(Random Forest,RF)算法,基于AWIFS、MODIS和SRTM遥感数据(空间分辨率分别为56 m,250 m和30m)及501个实测样点数据对陕西省及其不同地貌区的土壤有机质空间分布状况进行了预测,比较分析了不同分辨率、不同获取日期的遥感数据对土壤有机质预测结果的影响,通过随机百分比抽样对RF模型和普通克里金模型预测的土壤有机质的精度进行了比较分析,并对影响陕西省土壤有机质空间分布的因素进行了分析。取得的主要结果如下:1.依据RF模型的相对重要性因子排序和因子与土壤有机质的相关系数,筛选出陕西省参与土壤有机质预测的环境因素。包括高程、坡度、经度,纬度和不同遥感数据分辨率及成像时间获取的植被因子(NDVI、TVI、RSR和RVI)。2.陕西省土壤有机质含量在南北方向上差异较大,以陕西南部的秦岭地区和大巴山地区为最高,大部分土壤有机质含量大于25 g kg-1,黄土高原南部的土壤有机质含量处于中等水平,大部分在22-30 g kg-1之间,平原区的土壤有机质含量偏低,大部分低于30g kg-1,其中,关中平原的土壤有机质在4.43-16g kg-1之间,汉中低山丘陵地区的土壤有机质的含量大部分在9.48-30g kg-1,关中平原的东部局部低于10g kg-1,而黄土高原北部和风沙高原区的大部土壤有机质的含量最低,基本上在10g kg-1以下。3.遥感数据空间分辨率对土壤有机质的预测结果有一定的影响,在陕西省域尺度上基于AWIFS影像的预测效果要优于MODIS影像。而在定边县域尺度上基于TM影像的土壤有机质预测效果优于AWIFS影像。4.不同获取日期的遥感影像对土壤有机质预测的影响并不大,整体上,在平原地区,秋季获取的影像要比春季获取的影像预测的土壤有机质的值稍低。5.基于RF模型的土壤有机质预测精度在设定的不同抽样百分比条件下,比基于OK模型的土壤有机质预测精度高。同时,与已有的研究成果相比,本文的土壤有机质预测精度在期望值的范围内,独立验证数据集的平均误差大部分不超过3g kg-1,土壤有机质预测值与实测值的相关系数在0.7左右。6、在整个陕西省范围内,高程是影响土壤有机质预测的最重要因子,当影像的空间分辨率降低时,样点分布的地理经纬度和坡度(Slope)对土壤有机质预测的影响度上升,植被因子对土壤有机质的影响程度下降。
[Abstract]:Soil organic matter is one of the most representative attributes of soil and is an effective index to evaluate soil fertility and soil quality. Investigation and sampling is the main method to understand the status of soil organic matter in the region, but the prediction of soil organic matter only depends on the measured samples, which is time-consuming and laborious, and has the problem of large deviation in the local special terrain and geomorphology area. With the development of 3s technology, soil attribute prediction based on auxiliary environmental factors has been paid more and more attention in digital soil mapping. As a data source reflecting soil composition and vegetation growth, remote sensing data has been gradually applied to soil attribute prediction and played a unique advantage. In this paper, the spatial distribution of soil organic matter in Shaanxi Province and its different geomorphological regions was predicted by using random forest random forest fission (RFF) algorithm, based on the remote sensing data of AWIFS MODIS and SRTM (spatial resolution of 56mmm 250m and 30m respectively) and 501 measured samples. The effects of remote sensing data with different resolutions and different acquisition dates on soil organic matter prediction results were compared and analyzed. The accuracy of soil organic matter predicted by RF model and ordinary Kriging model was compared by random percentage sampling. The factors influencing the spatial distribution of soil organic matter in Shaanxi Province were analyzed. The main results are as follows: 1. Based on the ranking of relative importance factors of RF model and the correlation coefficient between factors and soil organic matter, the environmental factors involved in the prediction of soil organic matter in Shaanxi Province were selected. Vegetation factors, including elevation, slope, longitude, latitude and resolution of different remote sensing data and imaging time, are obtained by NDVIV TVI RSR and RVI. 2. The content of soil organic matter in Shaanxi Province varies greatly in the north and south directions, the Qinling region and the Dabashan region in the south of Shaanxi Province are the highest, the majority of the soil organic matter content is more than 25 g kg -1, and the content of soil organic matter in the southern part of the Loess Plateau is in the middle level. Most of the soil organic matter was between 22 and 30 g kg-1, and the content of soil organic matter was relatively low in the plain area, most of which was lower than 30 g kg -1. Among them, the soil organic matter in Guanzhong Plain ranged from 4.43-16 g kg-1. The content of soil organic matter was mostly 9.48-30g kg-1 in Hanzhong lowland and lower than 10g kg-1in the eastern part of Guanzhong Plain, while the content of soil organic matter was the lowest in the northern Loess Plateau and wind-sand plateau, basically below 10g kg-1. The spatial resolution of remote sensing data has a certain influence on the prediction results of soil organic matter, and the prediction effect based on AWIFS image is better than that of MODIS image in Shaanxi province. In Dingbian county scale, the prediction effect of soil organic matter based on TM image is better than that of AWIFS image. 4. Remote sensing images with different acquisition dates have little effect on soil organic matter prediction. On the whole, in plain areas, the values of soil organic matter obtained in autumn are slightly lower than those in spring. The prediction accuracy of soil organic matter based on RF model is higher than that based on OK model under different sampling percentages. At the same time, compared with the existing research results, the prediction accuracy of soil organic matter in this paper is within the range of expected value. The average error of independent validation data set is not more than 3 g kg -1, the correlation coefficient between soil organic matter prediction value and measured value is about 0.7. Height is the most important factor affecting soil organic matter prediction in the whole Shaanxi Province. When the spatial resolution of the image decreases, the influence of the geographic longitude, latitude and slope of the sample distribution on soil organic matter prediction increases, and the influence of vegetation factors on soil organic matter decreases.
【学位授予单位】：西北农林科技大学
【学位级别】：硕士
【学位授予年份】：2016
【分类号】：S153.621

【相似文献】