multinomial logistic regression

Digital Soil Mapping (DSM) encompasses a variety of methodologies that can yield precise spatial information about soil by establishing quantitative relationships between environmental covariates (predictors) and soil classes or properties. In this study, Artificial Neural Networks (ANNs), Decision Tree (DT), Multinomial Logistic Regression (MLR), and Random Forest (RF) algorithms were used to predict the soil map of downstream lands of Azad dam with an area of approximately 178.3 ha in the northwest of Sanandaj city in Kurdistan province. A random soil sampling method was used to determine the location and distribution of the 84 soil profiles in the study area. After recording soil morphological attributes, sampling of all horizons was conducted for required laboratory analysis. Afterward, the soil profiles were classified up to the family taxonomic level based on US classification system. Based on the soil taxonomy classification system, Inceptisols and Entisols order were observed by frequency, two Suborder, three Great groups, five Subgroups, and Family. To calculate the predictor variables, a digital elevation model (DEM) with a 10 m spatial resolution and Sentinel 2-B satellite images were used in the study area. To check the prediction accuracy of the models the Overall accuracy (OA), Kappa Index (K), and Brier Score (BS) were used. The best result was obtained by the ANN model (OA=0.65, K=0.53, and BS=0.16, respectively). The weakest predictions were found by DT model with OA, K, and BS of 0.38, 0.22, and 0.87, respectively.

Digital soil mapping (DSM), which uses machine learning (ML) algorithms and environmental covariates, is used worldwide for predicting soil properties and classes, due to being time-effective and cost-saving. This research was conducted to compare ML models and their efficiency in predicting the spatial distribution of soil subgroups in part of Zanjan Province. For this purpose, sampling was carried out through a regular pattern with 500 meters intervals, and 148 soil profile samples were randomly collected and classified. The soils of the region at the subgroup level were categorized in five classes, including Typic Calcixerepts, Typic Haploxerepts, Gypsic Haploxerepts, Typic Xerorthents, and Lithic Xerorthents. Environmental covariates included geomorphological and geological maps, digital elevation model (DEM), and remote sensing (RS), selected by principal component analysis (PCA) and expert knowledge approaches. Fifty-seven environmental variables were extracted as representatives of soil forming factors. Modeling of the soil-landscape relationship was performed using three algorithms, namely, multinomial logistic regression (MNLR), random forest (RF), and boosted regression tree (BRT) in Rstudio software. The results of evaluation metrics such as overall accuracy and Kappa index were 65% and 0.32 for the RF algorithm, 60% and 0.35 for the BRT, and 65% and 0.41 for the MNLR. Referring to the importance of environmental variables, results showed that the four factors of valley depth, LS-Factor, channel network distance, and multiple resolution ridge top flatness index (MRRTF) were the most important variables in MNLR algorithm. Also, the results of two statistics of user’s accuracy (UA) and producer’s accuracy (PA) showed that the MNLR model performed better in the spatial prediction of soil at subgroup level. Soil classes with higher frequency had higher accuracy. The results of the prediction accuracy of soil classes showed that the model prediction is more accurate for the more frequent types in the observations.

Effective and sustainable soil management requires knowledge about the spatial patterns of soil variation and soil surveys are important and useful sources of data that can be used. Prior knowledge about the spatial distribution of the soils is the first essential step for this aim but this requires the collection of large amounts of soil information. However, the conventional soil surveys are usually not useful for providing quantitative information about the spatial distribution of soil properties that are used in many environmental studies. Recently, by the rapid development of the computers and technology together with the availability of new types of remote sensing data and digital elevation models (DEMs), digital and quantitative approaches have been developed. These new techniques relies on finding the relationships between soil properties or classes and the auxiliary information that explain the soil forming factors or processes and finally predict soil patterns on the landscape. Different types of the machine learning approaches have been applied for digital soil mapping of soil classes, such as the logistic and multinomial logistic regressions, neural networks and classification trees. In reality, soils are physical outcomes of the interactions happening among the geology, climate, hydrology and geomorphic processes. Diversity is a way of measuring soil variation. Ibanez (9) first introduced ecological diversity indices as measures of diversity. Application of the diversity indices in soil science have considerably increased in recent years. Taxonomic diversity has been evaluated in the most previous researches whereas comparing the ability of different soil mapping approaches based on these indices was rarely considered. Therefore, the main objective of this study was to compare the ability of the conventional and digital soil maps to explain the soil variability using diversity indices in the Shahrekord plain of Chaharmahal-Va- Bakhtiari province.
The soils in the study area have been formed on Quaternary shale and foliated clayey limestone deposits. Irrigated crops such as wheat, barley and alfalfa are the main land uses in the area. According to the semi-detailed soil survey, 120 pedons with approximate distance of 750 m were excavated and described according to the “field book for describing and sampling soils”. Soil samples were taken from different genetic horizons and soil physicochemical properties were determined. Based on the pedons description and soil analytical data, pedons were classified according to the Soil Taxonomy (ST) up to subgroup level. Using aerial photo interpretation, geology map, google earth image and field observations primary soil map was created. With considering the taxonomic level, the representative pedons were determined and soil map was prepared. Multinomial logistic regression was used to predict soil classes at great group and subgroup levels. The map units that have the highest frequency were selected as indicator to calculate diversity indices in the conventional soil map at each taxonomic level. The selected map units were overlay to digital soil map and further diversity indices were calculated. Diversity indices including the Shannon’s diversity, evenness and richness index. In order to know whether the means of Shannon’s diversity for two approaches are significantly different, means comparison was done.
The results confirmed that the Shannon's diversity index was higher in the digital soil map than the conventional soil map for most soil map units. At great group and subgroup levels, a significant difference was observed for the Shannon's diversity index at 0.05 and 0.001 probability levels, respectively. Comparing the conventional and the digital soil maps showed the numbers of soil map units with significant difference regarding the Shannon's diversity index decreased from great group to the subgroup level. Although the conventional soil map did not show a good efficiency to explain the soil variability in this region considering more soil information to select the representative pedons at subgroup level in the conventional soil mapping could increase the ability of this approach.
A significant difference for the Shannon's diversity index between the conventional and the digital soil maps demonstrated that conventional soil mapping has not enough ability to explain the soil variability. It is recommended to test the effect of soil mapping approaches on explanation of the soil variability in other areas. Despite the deficiencies of traditional soil survey, it is still difficult to state about their replacement by digital methods.