Spatial modeling of land susceptibility to subsidence using machine learning models (case study: Kohdasht plain)

Introduction

Land subsidence is one of the biggest environmental threats worldwide, which has negative consequences on human and natural ecosystems. In this research, four machine learning methods, Bagged CART, GAM, Glmnet, and PLS, were used for spatial modeling of subsidence in Kohdasht plain in Lorestan province. For this purpose, first the most important factors affecting subsidence were investigated and finally 16 main factors were selected and their spatial layer was provided. In this research, the meta-heuristic algorithm of particle swarm optimization was used to select the main features and factors. Among the 16 input factors, 11 factors were selected as inputs for machine learning models. In the next step, subsidence points collected on the plain were used for subsidence modeling. The results of the models were evaluated using the ROC diagram and the area under the curve (AUC). The results showed that the used models have strong efficiency in subsidence modeling. The level under the curve for the PLS model was 0.981, which showed that this model is acceptable for preparing the map of land sensitivity to subsidence in the Kohdasht plain. The subsidence maps prepared by the models showed that 37.78% of the plain area in the PLS model is in the high risk class. The results of this research can be used as basic information to help planners and local officials in order to evaluate, plan, manage, sustainably use and protect the water resources of Kohdasht Plain in the future.

Materials and Methods

In this research, four machine learning methods, Bagged CART, GAM, Glmnet, and PLS, were used for spatial modeling of subsidence in Kohdasht plain in Lorestan province. For this purpose, first the most important factors affecting subsidence were investigated and finally 16 main factors were selected. These factors include topographical factors (height, percentage of slope and direction of slope), soilological factors (specific gravity of soil, cationic exchange capacity of soil, amount of clay, amount of silt, amount of sand, amount of coarse soil particles, amount of nitrogen and organic carbon of soil ), hydrological and hydrogeological factors (average annual drop of underground water, distance from production wells and distance from the river) and other factors (land use and lithology). And their spatial layer was provided. The second input related to the spatial modeling of machine learning models is the determination of subsidence points. 104 subsidence points were selected as modeling points. In the next step, 30% of the points for evaluation and 70% of the points for training the models were randomly separated. In order to evaluate the model by using the area under the curve (AUC) in the relative operator characteristic (ROC) curve, the prediction power of the model was checked and confirmed by using the nuclear cell surface index (SCAI). In the next step, the meta-heuristic algorithm of particle swarm optimization was used to select the main features and factors.



Results and Discussion

PSO method was used in this research. . From the total of 16 input factors, 11 are the most important factors, including height, slope direction and slope percentage, soil cation exchange capacity, amount of coarse soil particles, soil sand content, groundwater loss, distance from the river and distance from the well, land use and lithology. selected. After preparing the layers and initial information, 70% of the points were used to train the models and the subsidence sensitivity maps were prepared with four machine learning methods. After preparing the risk layers from the used models, 30 percent of the points were used for validation and evaluation. For this purpose, the validation points were placed on the subsidence risk layers and the value values of each layer were extracted. The results of the models were evaluated using the ROC chart and the AUC below it. The results showed that the used models had strong efficiency in subsidence modeling. The subsidence maps prepared by the models showed that 37.78% of the area of the plain in the PLS model and 21.83% of the area in the Glmnet model are in the high risk class. Also, examining the interpretability of the models showed that the factors of distance from the well, lithology, land use and coarse soil particles are the most important factors affecting the subsidence of the Kohdasht plain.



Conclusion

In the past several years, the use of machine learning models in the spatial modeling of environmental hazards is one of the attractive fields for research. In this research, the effectiveness of four machine learning methods Bagged CART, GAM, Glmnet and PLS in preparing land subsidence risk map in Kohdasht plain in Lorestan province has been evaluated. The results of the research showed that the machine learning models used along with remote sensing techniques and geographic information system provide powerful tools to assess the risk of land subsidence. The results of this research indicate the state of high subsidence potential in Kohdasht Plain. Based on the results of this research, machine learning models have great ability in spatial modeling, which is consistent with the results of Tien Bui et al. (2018), Oh et al. (2019), Zhou et al. (2019) and Sekkeravani et al. (2022). It is readable. The results of this research can be used as basic information to help planners and local officials in order to evaluate, plan, manage, sustainably use and protect the underground water resources of Kohdasht Plain. Considering the cost and time-consuming nature of soil tests and local visits, it is suggested to prepare a subsidence risk map using the method used in this research in other areas.