基于多源数据和Stacking-SHAP方法的山地丘陵区土地覆被分类

山地丘陵区地形复杂,地表辐射信号畸变严重,地物识别困难。为准确提取山区地物信息,结合多源异构数据,Stacking 集成学习和Shapley Addictive Explanation（SHAP）方法展开土地覆被分类研究。从Sentinel-1/2影像、气候数据、土壤数据和数字高程图中提取遥感、气候、土壤和地形四类特征变量,设计多种变量组合方案,结合Stacking算法,探讨不同类型变量在山区地物识别中的效用,并对比Stacking最佳方案与支持向量机（Support Vector Machine,SVM）、随机森林（Random Forest,RF）和极端梯度回归（eXtreme Gradient Boosting,XGBoost）算法的分类精度,评价Stacking方法在山区地物信息提取中的性能。同时,引入SHAP方法,量化Stacking模型中各特征变量的重要性。结果表明：在仅以遥感变量为基础方案时,山区土地覆被分类精度最低;在分别加入气候、土壤和地形变量后,总体精度、Kappa系数和F1分数均有所提高,其中旱地、水田和园地分类精度的提升幅度较大。基于Stacking算法结合所有类型特征变量的方案达到了最佳的分类精度,其总体精度、 Kappa系数和F1分数分别为96.61%、0.96和94.81%,分类精度优于相同特征下的SVM、 RF和XGBoost。SHAP方法可量化Stacking模型中特征变量的全局以及局部重要性,明确各变量对不同地物类型识别的相对贡献,为山区土地覆被分类的变量选择及优化提供有价值的信息。该研究为机器学习协助复杂景观地区土地覆被制图研究提供了技术支持和理论参考。     

Soils of the Verkhoyansk Range in the Area of Arkachan Plateau

Eurasian Soil Science - The vertical zonality of the soil cover of the Verkhoyansk mountainous soil province is discussed. The soils of the landscape-ecological profile in the area of the Arkachan...     

基于不同土壤分类系统的土壤景观格局分析——以我国西北黑河流域中游地区为例

对土壤景观格局进行的研究多是基于发生分类土壤图或通过参比转换得到的系统分类土壤图,尚无通过土壤系统分类调查直接得到的土壤图为基础进行相关研究的报道。本文依据目前可获得的我国西北黑河流域中游的20世纪80年代形成的发生分类土壤图和2012年及2013年通过系统调查采样形成的1︰50万系统分类土壤图,进行土壤景观格局分析对比。结果表明:无论系统分类还是发生分类,从类型水平来看,土壤的破碎化程度不高,被分割程度小、连通性高,土壤类型斑块形状偏简单;从景观水平来看,景观异质性较大,土壤类型数目较多,各土壤类型所占比例较均匀,土壤类型具有一定程度的积聚,土壤类型的连通度较高。与发生分类土壤图相比,系统分类的土壤类型斑块数较多,多样性指数和均匀度指数较高,蔓延度指数较低,说明在一定尺度和区域上,系统分类能更多地反映土壤类型空间上的差异,制图精度更高。     

Extrapolating regional soil landscapes from an existing soil map: Sampling intensity,validation procedures,and integration of spatial context

This paper aims to investigate the potential of using soil-landscape pattern extracted from a soil map to predict soil distribution at unvisited location. Recent machine learning advances used in previous studies showed that the knowledge embedded within soil units delineated by experts can be retrieved and explicitly formulated from environmental data layers However, the extent to which the models can yield valid prediction has been little studied. Our approach is based on a classification tree analysis which has underwent a recent statistics advance, namely, stochastic gradient boosting. We used an existing soil-landscape map to test our methodology. Explanatory variables included classical terrain factors (elevation, slope, curvature plan and profile, wetness index, etc.), various channels and combinations of channels from LANDSAT ETM imagery, land cover and lithology maps. Overall classification accuracy indexes were calculated under two validation schemes, either taken within the training area or from a separated validation area. We focused our study on the accuracy assessment and testing of two modelling parameters: sampling intensity and spatial context integration. First, we observed strong differences in accuracy between the training area and the extrapolated area. Second, sampling intensity, in proportion to the class extent, did not largely influence the classification accuracy. Spatial context integration by the use of a mean filtering algorithm on explanatory variables increased the Kappa index on the extrapolated area by more than ten points. The best accuracy measurements were obtained for a combination of the raw explanatory dataset with the filtered dataset representing regional trend. However, the predictive capacity of models remained quite low when extrapolated to an independent validation area. Nevertheless, this study offers encouragement for the success of extrapolating soil patterns from existing soil maps to fill the gaps in present soil map coverage and to increase efficiency of ongoing soil survey.     

基于多源环境变量的橡胶园土壤管理分区

为了提高大尺度范围内橡胶园土壤管理的针对性,以海南省国营八一农场橡胶园为研究对象,以地形因子(海拔、坡度和坡向)、成土母质、气候变量(平均降雨量和平均气温)和植被指数为数据源,对橡胶园进行土壤管理分区。利用模糊C均值聚类法进行分区,以模糊性能指数(fuzzy performance index,FPI)和归一化分类熵(normalized classification entropy,NCE)作为判断最佳分区数的标准,并通过单因素方差分析和分区前后土壤属性以及环境变量变异系数对比对分区结果进行评价和验证。研究结果表明,橡胶园管理分区的最佳数目为3个。不同管理分区之间土壤属性(pH值、有机质、全氮、有效磷、速效钾、有效硫、交换性钙、交换性镁、有效铜、有效铁和有效锰)和环境变量(高程、坡度、降雨量、平均温和归一化植被指数)的差异性都达到了极显著水平(P0.01),同时,3个分区中土壤属性和环境变量变异系数的均值比分区前明显下降。这就验证了在大尺度范围内,利用较易获取的多源环境变量进行橡胶园土壤管理分区是可行的,同时依据不同分区的特点制定了相应的土壤管理措施,提高了大范围区域内橡胶园土壤管理的针对性。     

Relationships of some environmental variables to the aggregation patterns of soil microarthropod populations in forests

Soil Collembola populations exhibit a non-random distribution pattern, which has been associated with a number of environmental variables. The purpose of this study is a) to investigate the patterns of aggregation and the relationship of the population density of selected soil Collembola species to environmental variables and b) to test whether numerical abundance and biomass as measures of population density lead to consistent results regarding these aspects of their ecology. A gradient of soil organic matter content and a gradient of soil pH, both occurring within limited space, were studied in two different sites. Taylor's b was used as an aggregation index. Two multiple regression models were estimated for each dominant species in each study site linking numerical abundance or biomass to environmental variables. The species studied exhibited different degrees of aggregation. These differences matched closely the differences in the strength of the association between population density and soil organic matter or pH, depending on the site. The regression models estimated for the biomass of species for which accurate methods of biomass estimation are available, exhibited a better fit than the models estimated for the numerical abundance of the same population.     

Multivariate approach to evaluate the penetrometer resistance in different tillage systems

At present, the penetrometer is the most widely used instrument for assessing in situ soil strength, one of the extrinsic factors affecting plant growth and crop productivity. In this paper we propose a method that discriminates penetrometer resistance due to different soil treatments, by means of Principal Component (PC) analysis. We hypothesized and demonstrated that penetrometer resistance values measured at different soil depths are correlated among themselves (multicollinearity). Considering measurements at each depth as different variables, PC analysis restructured data sets containing these correlated variables into a smaller number of components, whose scores were utilized in univariate analysis of variance (ANOVA) to test differences among imposed soil treatments. We applied the procedure to penetrometer resistance values measured by means of a hand-held cone penetrometer in two long-term experiments conducted in southern Italy, on durum wheat (Triticum durum Desf.) under continuous cropping. In the first trial, four different soil tillage treatments were compared; in the second, two different tillages and two residue management systems were examined. In both trials, PC analysis reduced the original 14 depths of measurements into only 4 PC's, based on correlations of their resistance values, explaining more than 80% of the total variation. Furthermore, ANOVA applied to the scores of each PC, clearly indicated treatment effects on soil strength.

The proposed method has thus allowed assemblage a posteriori of penetration resistance data into only a few significative intervals, using correlations among the measurements made at the different depths. This way, the possible resistance differences due to tillage and/or management treatments have been more easily and more unambiguously showed.     

宁夏固原县侵蚀土壤的数值分析研究

数学作为一种手段,应用于侵蚀土壤的研究,必定对这一研究有所推动。本文用宁夏固原县侵蚀土壤的观测资料,采用多元分析方法,借助电子计算机,对侵蚀土壤的数值分类、不同侵蚀土壤类型与侵蚀因子的关系进行研究。     

Soil classification provides a poor indicator of carbon turnover rates in soil

Most soil surveys are based on soil geomorphic, physical and chemical properties, while many classifications are based on morphological properties in soil profile. Typically, microbial properties of the soil (e.g. biomass and functional diversity) or soil biological quality indicators (SBQIs) are not directly considered in soil taxonomic keys, yet soil classification schemes are often used to infer soil biological function relating to policy (e.g. soil pollution attenuation, climate change mitigation). To critically address this, our aim was to assess whether rates of carbon turnover in a diverse range of UK soils (n > 500) could effectively be described and sub-divided according to broadly defined soil groups by conventional soil classification schemes. Carbon turnover in each soil over a 90 d period was assessed by monitoring the mineralisation of either a labile (¹⁴C-labelled artificial root exudates) or more recalcitrant C source (¹⁴C-labelled plant leaves) in soil held at field capacity at 10 °C. A double exponential first order kinetic model was then fitted to the mineralisation profile for each individual substrate and soil. ANOVA of the modelled rate constants and pool sizes revealed significant differences between soil groups; however, these differences were small regardless of substrate type. Principle component and cluster analysis further separated some soil groups; however, the definition of the class limits remained ambiguous. Exclusive reference values for each soil group could not be established since the model parameter ranges greatly overlapped. We conclude that conventional soil classification provides a poor predictor of C residence time in soil, at least over short time periods. We ascribe this lack of observed difference to the high degree of microbial functional redundancy in soil, the strong influence of environmental factors and the uncertainties inherent in the use of short term biological assays to represent pedogenic processes which have taken ca. 10,000 y to become manifest.     

城市土壤特征土层的数值分类

Pedogenetic soil horizons are one of the fundamental building blocks of modern soil classification; however, in soils of urban areas which are often strongly disturbed by human activities, horizons are difficult to distinguish but substitutive morphological layers may be identified. To identify the characteristic soil layers in an urban environment, 224 soil layers of 36 in-situ pedons were examined and described in urban and suburban Nanjing, and 27 variables were extracted for multivariate analysis. Three groups and six subdivisions were identified by TwoStep cluster analysis combined with hierarchical cluster analysis based on factor scores. Soil forming factors and soil forming processes could be interpreted from the principal component analysis （PCA） of variables, cluster analysis of soil layers, and discriminant analysis of soil layer groups and their subdivisions. Parent materials, moisture regimes, organic matter accumulation, and especially nutrient accumulation were the main causes of characteristic soil layer formations. The numerical approaches used in this study were useful tools for characteristic soil layer identification of urban soils.     

利用侵蚀模型普查黄土高原土壤侵蚀状况

土壤侵蚀普查对于土地资源保护和自然灾害防治具有重要意义。为了测试抽样方法和土壤侵蚀模型在土壤侵蚀普查中的适用性,该文以陕西吴起县为试点,采用1%均匀抽样方法,调查39个抽样单元的土壤侵蚀影响因子,使用中国侵蚀预报模型CSLE(Chinese soilloss equation)估算土壤侵蚀模数,并与基于遥感数据的水蚀分级分类方法进行比较。两种方法估算的全县平均土壤侵蚀模数分别为4571和5504t/(km2a),但不同分级侵蚀强度的面积和空间分布存在较大差异。抽样方法在土地利用与覆盖、水土保持措施及土壤特性方面获得的信息量大于遥感方法,同时对于区域具有很好的代表性;使用模型估算土壤侵蚀考虑的影响因子与分级方法相比,还包括了土壤可蚀性、坡长因子以及水土保持措施因子等,由此计算的土壤侵蚀模数和强度具有更高的可信度。因此,虽然基于抽样方法和土壤侵蚀模型的土壤侵蚀普查方法也存在一定的问题,但与土壤侵蚀分类分级方法相比具有明显的优越性。     

Coping with imbalanced data problem in digital mapping of soil classes

An unsolved problem in the digital mapping of categorical soil variables and soil types is the imbalanced number of observations, which leads to reduced accuracy and the loss of the minority class (the class with a significantly lower number of observations compared to other classes) in the final map. So far, synthetic over- and under-sampling techniques have been explored in soil science; however, more efficient approaches that do not have the drawbacks of these techniques and guarantee retention of the minority classes in the produced map are essentially required. Such approaches suggested in the present study for digital mapping of soil classes include machine learning models of ensemble gradient boosting, cost-sensitive learning and one-class classification (OCC) of the minority class combined with multi-class classification. In this regard, extreme gradient boosting (XGB) as an ensemble gradient learner, a cost-sensitive decision tree (CSDT) within the C5.0 algorithm, and a one-class support vector machine combined with multi-class classification (OCCM) were investigated to map eight soil great groups with a naturally imbalanced frequency of observations in northwest Iran. A total of 453 profile data points were used for mapping the soil great groups of the study area. A data split was done manually for each class separately, which resulted in an overall 70% of the data for calibration and 30% for validation. The bootstrapping approach of calibration (with 10 runs) was performed to produce multiple maps for each model. The 10 bootstraps were evaluated against the hold-out validation dataset. The average values of accuracy measures, including Kappa (K), overall accuracy (OA), producer's accuracy (PA) and user's accuracy (UA), were explored. In addition, the results of this study were compared with a previous study in the same area, in which resampling techniques were used to deal with imbalanced data for digital soil class mapping. The findings show that all three suggested methods can deal well with the imbalanced classification problem, with OCCM showing the highest K (= 0.76) and OA (= 82) in the validation stage. Also, this model can guarantee the retention of the minority classes in the final map. Comparing the present approaches with the previous study approach demonstrates that the three newly suggested methods can remarkably increase both overall and individual class accuracy for mapping.     

环境敏感变量优选及机器学习算法预测绿洲土壤盐分

基于机器学习预测干旱区(如新疆)土壤盐分的研究目前较少涉及且敏感变量的筛选还需深入探讨。该研究比较5种机器学习算法(套索算法,The Least Absolute Shrinkage and Selection Operator-LASSO;多元自适应回归样条函数,Multiple Adaptive Regression Splines-MARS;分类与回归树,Classification and Regression Trees-CART;随机森林,Random Forest-RF;随机梯度增进算法,Stochastic Gradient Treeboost-SGT)在3个不同地理区域(奇台绿洲,渭-库绿洲和于田绿洲)的性能表现;参与的变量被分为6组:波段,植被相关变量集,土壤相关变量集,数字高程模型(digital elevation model,DEM)衍生变量集,全变量组,优选变量组(全变量组经过算法筛选后的变量集合)。通过算法筛选,以示不同研究区的盐度敏感变量。同时借助以上述6组结果评判算法的性能。结果表明:综合分析6个变量组的R2和RMSE,预测精度排名如下:优选变量组植被指数变量组土壤相关变量组波段DEM衍生变量组。由于结果不稳定,全变量组未参与排名。在所有变量中,植被指数(EEVI,ENDVI,EVI2,CSRI,GDVI)和土壤盐度指数(SIT,SI2和SAIO)与土壤盐度相关性高于其他变量。综合评价以上5种算法,Lasso和MARS的预测结果出现极端异常值,但其预测结果能基本呈现土壤盐分空间分布格局。CART的结果能清晰分辨灌区和非灌区土壤盐分的分布态势,但二者内部并无太多变化且稳定性较差。RF和SGT的结果显示,二者在3个绿洲的土壤盐分值域范围和土壤盐分空间分布格局相似,纹理信息相对其他3个算法更为丰富。更为重要的是,算法在各个地区的结果都较为稳定。二者相比,SGT验证精度相对最高,其次为RF。     

Fuzzy classification methods for determining land suitability from soil profile observations and topography

Because conventional Boolean retrieval of soil survey data and logical models for assessing land suitability treat both spatial units and attribute value ranges as exactly specifiable quantities, they ignore the continuous nature of soil and landscape variation and uncertainties in measurement which can result in the misclassification of sites that just fail to match strictly defined requirements. This paper uses fuzzy classification to determine land suitability from (i) multivariate point observations of soil attributes, (ii) topographically controlled site drainage conditions, and (iii) minimum contiguous areas, and compares the results obtained with conventional Boolean methods. The methods are illustrated using data from the Alberta Agricultural Department experimental farm at Lacombe in Alberta, Canada. Data on site elevation and soil chemical and physical properties measured at 154 soil profiles were interpolated by ordinary block kriging to 15 m × 15 m cells on a 50 × 50 grid. The soil property data for each cell were classified by Boolean and fuzzy methods. The digital elevation model created by interpolating the elevation data was used to determine the surface drainage network and map it in terms of the numbers of cells draining through each cell on the grid. This map was reclassified to yield Boolean and fuzzy maps of surface wetness which were then intersected with the soil profile classes. The resulting classification maps were examined for contiguity to locate areas where a block of minimum size (45m × 45m) could be located successfully. In this study Boolean methods reject larger numbers of cells than fuzzy classification, and select cells that are insufficiently contiguous to meet the aims of the land classification. Fuzzy methods produce contiguous areas and reject less information at all stages of the analyses than Boolean methods. They are much better than Boolean methods for classification of continuous variation, such as the results of the drainage analysis.     

Soil-landform relationships at two localities in Papua New Guinea

Two test areas, each of several hectares, in denudational terrain in the lowlands and highlands of Papua New Guinea were intensively studied to establish the association between soil conditions and landforms. Whether soils were allocated either to the great soil groups of a classification based on the 7th Approximation or to FAO soil units, an extremely wide range of soils occurred in each of the land units that could be distinguished on the basis of landform. Defensible statements concerning soil distribution were achieved only by limiting the categories of soil to the level of soil orders and by restricting the number of land units to three or four. Even so, the statistical significance of the relationships was very low.These results indicate that current techniques and sampling densities used in photointerpretative land resource surveys in terrains typical of Papua New Guinea can produce only very generalized information on soil distribution. To obtain reliable predictions of soil information at the level commonly implied in such surveys, further intensive studies on the relationships between attributes of soil, landform and vegetation are required.     

Comparing mapping approaches at subcatchment scale in northern Thailand with emphasis on the Maximum Likelihood approach

This study compares different soil mapping approaches in three different petrographic areas in order to test their suitability for regional mapping in northern Thailand. Sampling was based on transects or grid-based randomization. Maps were created based on expert knowledge (eye fitting) or using Classification Tree (CART algorithm) or the Maximum Likelihood approach. In addition, local knowledge-based-soil maps were created. Validation was performed using soil reference maps and independent sampling points. The mapping approaches based on transects and grid-based randomization showed a very high correspondence with the respective reference soil map and a very high degree of matching with independent sampling points. Both methods are best suited for sub-watershed scale. Mapping larger areas is difficult due to the inaccessibility of the mountainous regions. The soil maps based on Maximum Likelihood showed a high correlation with the respective reference soil maps and the individual sampling points. Maximum Likelihood maps and Classification Tree maps showed similar levels of accuracy. The Maximum Likelihood approach is applicable to upscaling procedures; therefore, a calibration area is required which represents the target area. Local knowledge-based-soil mapping is very cheap and fast, but is restricted to village areas where classification often varies even within a village. Despite this, local knowledge is very useful for soil reconnaissance surveys, as well as to acquire an overview of the major distribution of soils and their properties. Upscaling of local knowledge due to its inherent inconsistency is not realistic.     

Increasing the relevance of scientific information in hillside environments through understanding of local soil management in a small watershed of the Colombian Andes

Abstract. This article explores the question of how scientific information can improve local agronomic management using concepts of uncertainty classification and uncertainty management. Information and data on local management of soil fertility based on a local classification system of soil quality were collected from a small watershed in Cauca (Colombia). The analyses suggest that farmers hold local knowledge about soils at two levels. The first is based on empirical observations and refers to local knowledge about soils and landscape, which shows that the classes identified in the local soil quality classification are consistent with results obtained using measured soil parameters. At the second level, farmers have some awareness of ecological processes and the appropriate use of relationships between key soil characteristics and management options. It is argued that local knowledge is not sufficient to cope with uncertainty introduced by a rapidly changing agriculture, including, for example, increasing land pressure, unpredictable market forces and climate change. We have suggested how scientific knowledge can contribute to the solution, based on an analysis that relates Cohen's ( Heuristic reasoning about uncertainty: an artificial intelligence approach . Pitman London, 1985) and Rowe's ( Risk Analysis 14, 743–750, 1994) uncertainty concepts to local knowledge.     

Detecting zones of abrupt change in soil data, with an application to an agricultural field

We propose a new method for estimating and testing the zones where a variable has discontinuities or sharp changes in the mean. Such zones are called Zones of Abrupt Change (ZACs). Our method is based on the statistical properties of the estimated gradient of the variable. The local gradient is first interpolated by kriging. Then we test whether the estimated local gradient exceeds some critical threshold computed under the null hypothesis of a constant mean. The locations where the local test is rejected define the potential ZACs, which are then tested globally. Using this method, we analysed soil data from an agricultural field. The analysis of the main soil components of the ploughed layer (clay, silt and sand particles and calcium carbonate content) reveals the structural variations in the field, linked to boundaries between soil types. Its application to non‐permanent variables (soil water and mineral nitrogen content of the soil profile to 120 cm taken at several dates) shows that water content has the same ZACs for all dates, whereas mineral nitrogen has none.