5:15 PM - 6:30 PM
[AGE27-P08] Neural network estimation of soil organic matter contents and its relation to other soil components.
Keywords:Neural Network, Soil Organic Carbon, Hydraulic properties, pedo-transfer function
Soil is the largest carbon pool on land and plays an important role in supplying nutrients to animals and plants and improving the physical properties of soil. Currently, various SOC dynamic prediction models have been proposed, but most of them have been developed in Europe and the United States, and the application examples are biased toward temperate field soils. Until now, the human side has narrowed down the parameters and pursued the functional type, but if the optimum relationship can be derived from the data itself, we might clarify the phenomenon in the natural world. The objectives this study were to create a SOC prediction model using a neural network (NN) model, and to consider the relationship between SOC and other soil components in the target area.
Soil data for Europe, Southeast Asia, Australia, North America and South America were extracted from the Harmonized world soil database v1.2. In this study, in addition to using all features, model-based feature selection (Lasso (alpha = 1), Lasso (alpha = 10), Random Forest, and recursive feature selection (Random Forest parameters 8 to 1) are used as feature selection methods. In addition, 7 types of regularization strength (Penalty lambda=10, 1, 0.1, 0.01, 0.001, 0.0001, 0) in the objective function were prepared, models were created 5 times each, and accuracy verification was performed by cross-validation. 100 models were created by combining parameters with high prediction accuracy, and then the selected features were evaluated. In addition, hydraulic parameters were derived from the HWSD particle size data using ROSETTA, which is a Pedotransfer Function model, and the selected features were evaluated.
The features commonly selected in the models were silt content, dry density, cation exchange capacity, and base saturation. The development of soil pore structure and the adsorption of high-valence ions are cited as factors in the amount of organic matter, which are consistent with conventional knowledge. In addition, when hydraulic parameters are added to the features, they are selected as the features at each target site, thus, we could say the infiltration or water retention properties is a factor related to organic matter conservation.
Soil data for Europe, Southeast Asia, Australia, North America and South America were extracted from the Harmonized world soil database v1.2. In this study, in addition to using all features, model-based feature selection (Lasso (alpha = 1), Lasso (alpha = 10), Random Forest, and recursive feature selection (Random Forest parameters 8 to 1) are used as feature selection methods. In addition, 7 types of regularization strength (Penalty lambda=10, 1, 0.1, 0.01, 0.001, 0.0001, 0) in the objective function were prepared, models were created 5 times each, and accuracy verification was performed by cross-validation. 100 models were created by combining parameters with high prediction accuracy, and then the selected features were evaluated. In addition, hydraulic parameters were derived from the HWSD particle size data using ROSETTA, which is a Pedotransfer Function model, and the selected features were evaluated.
The features commonly selected in the models were silt content, dry density, cation exchange capacity, and base saturation. The development of soil pore structure and the adsorption of high-valence ions are cited as factors in the amount of organic matter, which are consistent with conventional knowledge. In addition, when hydraulic parameters are added to the features, they are selected as the features at each target site, thus, we could say the infiltration or water retention properties is a factor related to organic matter conservation.