An overview and comparison of machine-learning techniques for classification purposes in digital soil mapping

Heung, Brandon; Ho, Hung Chak; Zhang, Jin; Knudby, Anders; Bulmer, Chuck; Schmidt, Margaret G.

doi:10.1016/j.geoderma.2015.11.014

Cited by 402 publications

(224 citation statements)

References 84 publications

Supporting

Mentioning

216

Contrasting

Unclassified

Order By: Relevance

“…The tested methods included Partial Least Square Regression (PLSR) [63], Support Vector Machine Regression (SVMR) [64][65][66] (using different kernel functions-linear, polynomic and radial), random forest (RF) [67] and artificial neural network (ANN) [68]. More information can be found in Heung et al [69], Vasques et al [70] or Gholizadeh et al [27].…”

Section: Multivariate Techniquesmentioning

confidence: 99%

Assessment of Soil Degradation by Erosion Based on Analysis of Soil Properties Using Aerial Hyperspectral Images and Ancillary Data, Czech Republic

2017

View full text Add to dashboard Cite

Abstract:The assessment of the soil redistribution and real long-term soil degradation due to erosion on agriculture land is still insufficient in spite of being essential for soil conservation policy. Imaging spectroscopy has been recognized as a suitable tool for soil erosion assessment in recent years. In our study, we bring an approach for assessment of soil degradation by erosion by means of determining soil erosion classes representing soils differently influenced by erosion impact. The adopted methods include extensive field sampling, laboratory analysis, predictive modelling of selected soil surface properties using aerial hyperspectral data and the digital elevation model and fuzzy classification. Different multivariate regression techniques (Partial Least Square, Support Vector Machine, Random forest and Artificial neural network) were applied in the predictive modelling of soil properties. The properties with satisfying performance (R 2 > 0.5) were used as input data in erosion classes determination by fuzzy C-means classification method. The study was performed at four study sites about 1 km 2 large representing the most extensive soil units of the agricultural land in the Czech Republic (Chernozems and Luvisols on loess and Cambisols and Stagnosols on crystalline rocks). The influence of site-specific conditions on prediction of soil properties and classification of erosion classes was assessed. The prediction accuracy (R 2 ) of the best performing models predicting the soil properties varies in range 0.8-0.91 for soil organic carbon content, 0.21-0.67 for sand content, 0.4-0.92 for silt content, 0.38-0.89 for clay content, 0.73-089 for Fe ox , 0.59-0.78 for F ed and 0.82 for CaCO 3 . The performance and suitability of different properties for erosion classes' classification are highly variable at the study sites. Soil organic carbon was the most frequently used as the erosion classes' predictor, while the textural classes showed lower applicability. The presented approach was successfully applied in Chernozem and Luvisol loess regions where the erosion classes were assessed with a good overall accuracy (82% and 67%, respectively). The model performance in two Cambisol/Stagnosol regions was rather poor (51%-52%). The results showed that the presented method can be directly and with a good performance applied in pedologically and geologically homogeneous areas. The sites with heterogeneous structure of the soil cover and parent material will require more precise local-fitted models and use of further auxiliary information such as terrain or geological data. The future application of presented approach at a regional scale promises to produce valuable data on actual soil degradation by erosion usable for soil conservation policy purposes.

show abstract

Section: Multivariate Techniquesmentioning

confidence: 99%

Assessment of Soil Degradation by Erosion Based on Analysis of Soil Properties Using Aerial Hyperspectral Images and Ancillary Data, Czech Republic

2017

View full text Add to dashboard Cite

show abstract

“…Big data is practically reshaping all business sectors [1], as it is a great source of advancement and economic value, thus analyzing Big Data leads to better insights and new understanding in assisting of different sectors for better decision making. Machine learning is a technique used to be known as big data that assists to get deep insights [2], defined as the process of discovering the relationships between predictor and response variables using computerbased statistical approaches [3]. There are different kinds of statistical approach or methods are used in machine learning.…”

Section: Introductionmentioning

confidence: 99%

Impact of Different Data Types on Classifier Performance of Random Forest, Naïve Bayes, and K-Nearest Neighbors Algorithms

Singh¹,

Halgamuge²,

Lakshmiganthan³

2017

ijacsa

View full text Add to dashboard Cite

Abstract-This study aims to evaluate impact of three different data types (Text only, Numeric Only and Text + Numeric) on classifier performance (Random Forest, k-Nearest Neighbor (kNN) and Naïve Bayes (NB) algorithms). The classification problems in this study are explored in terms of mean accuracy and the effects of varying algorithm parameters over different types of datasets. This content analysis has been examined through eight different datasets taken from UCI to train models for all three algorithms. The results obtained from this study clearly show that RF and kNN outperform NB. Furthermore, kNN and RF perform relatively the same in terms of mean accuracy nonetheless kNN takes less time to train a model. The changing numbers of attributes in datasets have no effect on Random Forest, whereas Naïve Bayes mean accuracy fluctuates up and down that leads to a lower mean accuracy, whereas, kNN mean accuracy increases and ends with higher accuracy. Additionally, changing number of trees has no significant effects on mean accuracy of the Random forest, however, the time to train the model has increased greatly. Random Forest and k-Nearest Neighbor are proved to be the best classifiers for any type of dataset. Thus, Naïve Bayes can outperform other two algorithms if the feature variables are in a problem space and are independent. Besides Random forests, it takes highest computational time and Naïve Bayes takes lowest. The k-Nearest Neighbor requires finding an optimal number of k for improved performance at the cost of computation time. Similarly, changing the number of attributes that effect Naïve Bayes and k-Nearest Neighbor performance nevertheless not the Random forest. This study can be extended by researchers who use the parametric method to analyze results.

show abstract

“…A 30-m circular buffer was created in each pedon, similar to that performed by Sarmento et al (2014), , Heung et al (2016), and Vincent et al (2018). This buffer is created for the purpose of disregarding the precision error involved in the GPS equipment used to collect the geographic coordinates of the pedons, which may be more than 20 meters in some cases, depending on the atmospheric conditions and the equipment used.…”

Section: Map Accuracy Assessmentmentioning

confidence: 99%

Spatial Disaggregation of Multi-Component Soil Map Units Using Legacy Data and a Tree-Based Algorithm in Southern Brazil

Machado

Giasson

Campos

et al. 2018

Rev. Bras. Ciênc. Solo

View full text Add to dashboard Cite

Soil surveys often contain multi-component map units comprising two or more soil classes, whose spatial distribution within the map unit is not represented. Digital Soil Mapping tools supported by information from soil surveys make it possible to predict where these classes are located. The aim of this study was to develop a methodology to increase the detail of conventional soil maps by means of spatial disaggregation of multi-component map units and to predict the spatial location of the derived soil classes. Three digital maps of terrain variables -slope, landforms, and topographic wetness indexwere correlated with the soil map and 72 georeferenced profiles from the Porto Alegre soil survey. Explicit rules that expressed regional soil-landscape relationships were formulated based on the resulting combinations. These rules were used to select typical areas of occurrence of each soil class and to train a decision tree model to predict the occurrence of individualized soil classes. Validation of the soil map predictions was conducted by comparison with available soil profiles. The soil map produced showed high agreement (80.5 % accuracy) with the soil classes observed in the soil profiles; Ultisols and Lithic Udorthents were predicted with greater accuracy. The soil variables selected in this study were suitable to represent the soil-landscape relationships, suggesting potential use in future studies. This approach developed a more detailed soil map relevant to current demands for soil information and has potential to be replicated in other areas in which data availability is similar.

show abstract

An overview and comparison of machine-learning techniques for classification purposes in digital soil mapping

Cited by 402 publications

References 84 publications

Assessment of Soil Degradation by Erosion Based on Analysis of Soil Properties Using Aerial Hyperspectral Images and Ancillary Data, Czech Republic

Assessment of Soil Degradation by Erosion Based on Analysis of Soil Properties Using Aerial Hyperspectral Images and Ancillary Data, Czech Republic

Impact of Different Data Types on Classifier Performance of Random Forest, Naïve Bayes, and K-Nearest Neighbors Algorithms

Spatial Disaggregation of Multi-Component Soil Map Units Using Legacy Data and a Tree-Based Algorithm in Southern Brazil

Contact Info

Product

Resources

About