Gorgan University Of Agricultural Sciences Journal of Water and Soil Conservation 2322-2069 24 1 2017 03 21 Application of Random Forest method for predicting soil classes in low relief lands (case study: Hirmand county) Application of Random Forest method for predicting soil classes in low relief lands (case study: Hirmand county) 67 84 3577 10.22069/jwfst.2017.12396.2700 FA Khalililah Mirak Zehi 0009-0005-8306-0440 Ali Shahriari Mohammd Reza Pahlevanrad Abolfazl Bameri Journal Article 2016 12 22 Abstract<br />Background and Objectives: Base of soil information for environmental modeling is soil survey and mapping as a way to determine soil distribution patterns, describe and display it to understood and interpreted for different users. Digital soil mapping creates link between classes or soil characteristics and environmental factors affected soil formation and development by using mathematical models which can provide more precise and accurate soil maps and reducing cost of soil survey and mapping projects. This study was done to mapping soil great groups and subgroups by using Random Forest technique in the Hirmand county lands in Sistan plain.<br />Materials and Methods: In this study 108 soil profiles were dug on about 60.000 hectares of Hirmand county lands. Sixteen environmental variables were used as estimator for soil mapping including land properties, salinity and vegetation index. After classification of soil profiles to great groups and subgroups, soil classes map provided by using random forest (RF) method. It should be mentioned 80 percent of data was used for model training and 20 percent for independent validation. <br />Results: Pedological studies showed soils that formed in Sistan plain haven’t high development and most of them are Entisol and Aridisol. Most soil profiles classified as Torrifluvents on great groups and Typic Torrifluvents as subgroups. Also the result of RF showed the lowest estimation error of out of bag (OOB) samples in soil great groups and subgroups were 43.53 and 50.59 respectively. Independent validation results showed the best accuracy obtained for soil great groups and subgroups were 48 and 53 percent respectively. Grooves depth, convergence index, channel network between and salinity in soil great groups and grooves depth, elevation and catchment area in soil subgroups were the most important environmental variables to estimate soil classes.<br />Conclusion: The results showed most soils are young in the low relief lands in arid regions and these regions have also low soil diversity. Soil digital mapping and random forest technique could be useful for soil classes prediction and soil mapping in this kind of regions. <br />Results: Pedological studies showed soils that formed in Sistan plain haven’t high development and most of them are Entisol and Aridisol. Most soil profiles classified as Torrifluvents on great groups and Typic Torrifluvents as subgroups. Also the result of RF showed the lowest estimation error of out of bag (OOB) samples in soil great groups and subgroups were 43.53 and 50.59 respectively. Independent validation results showed the best accuracy obtained for soil great groups and subgroups were 48 and 53 percent respectively. Grooves depth, convergence index, channel network between and salinity in soil great groups and grooves depth, elevation and catchment area in soil subgroups were the most important environmental variables to estimate soil classes.<br />Conclusion: The results showed most soils are young in the low relief lands in arid regions and these regions have also low soil diversity. Soil digital mapping and random forest technique could be useful for soil classes prediction and soil mapping in this kind of regions. <br />Keywords: Soil digital mapping, Random forest technique, Map accuracy, Arid regions, Sistan plain Abstract<br />Background and Objectives: Base of soil information for environmental modeling is soil survey and mapping as a way to determine soil distribution patterns, describe and display it to understood and interpreted for different users. Digital soil mapping creates link between classes or soil characteristics and environmental factors affected soil formation and development by using mathematical models which can provide more precise and accurate soil maps and reducing cost of soil survey and mapping projects. This study was done to mapping soil great groups and subgroups by using Random Forest technique in the Hirmand county lands in Sistan plain.<br />Materials and Methods: In this study 108 soil profiles were dug on about 60.000 hectares of Hirmand county lands. Sixteen environmental variables were used as estimator for soil mapping including land properties, salinity and vegetation index. After classification of soil profiles to great groups and subgroups, soil classes map provided by using random forest (RF) method. It should be mentioned 80 percent of data was used for model training and 20 percent for independent validation. <br />Results: Pedological studies showed soils that formed in Sistan plain haven’t high development and most of them are Entisol and Aridisol. Most soil profiles classified as Torrifluvents on great groups and Typic Torrifluvents as subgroups. Also the result of RF showed the lowest estimation error of out of bag (OOB) samples in soil great groups and subgroups were 43.53 and 50.59 respectively. Independent validation results showed the best accuracy obtained for soil great groups and subgroups were 48 and 53 percent respectively. Grooves depth, convergence index, channel network between and salinity in soil great groups and grooves depth, elevation and catchment area in soil subgroups were the most important environmental variables to estimate soil classes.<br />Conclusion: The results showed most soils are young in the low relief lands in arid regions and these regions have also low soil diversity. Soil digital mapping and random forest technique could be useful for soil classes prediction and soil mapping in this kind of regions. <br />Results: Pedological studies showed soils that formed in Sistan plain haven’t high development and most of them are Entisol and Aridisol. Most soil profiles classified as Torrifluvents on great groups and Typic Torrifluvents as subgroups. Also the result of RF showed the lowest estimation error of out of bag (OOB) samples in soil great groups and subgroups were 43.53 and 50.59 respectively. Independent validation results showed the best accuracy obtained for soil great groups and subgroups were 48 and 53 percent respectively. Grooves depth, convergence index, channel network between and salinity in soil great groups and grooves depth, elevation and catchment area in soil subgroups were the most important environmental variables to estimate soil classes.<br />Conclusion: The results showed most soils are young in the low relief lands in arid regions and these regions have also low soil diversity. Soil digital mapping and random forest technique could be useful for soil classes prediction and soil mapping in this kind of regions. <br />Keywords: Soil digital mapping, Random forest technique, Map accuracy, Arid regions, Sistan plain https://jwsc.gau.ac.ir/article_3577_c7851b1a49d89163d15a9c96321ac348.pdf