Search

Predictive mapping of environment is an important means for environment assessment and management. The selection of predictor variables (or environmental covariates) is the first and key step in predictive mapping. A number of machine learning and statistical models have been developed to select what and how many environmental covariates in a wide range of predictive mapping. Nevertheless, those models require a large amount of field data for model training and calibration, which can be problematic in applying to the areas with no or very limited field data available. To overcome the shortcoming, this paper proposes the most similar case method for selecting environmental covariates for predictive mapping. First, we describe the basic idea and the development procedures of the most similar case method; second, as an experimental test, we employ the proposed method to select the topographic covariates for inputting to the predictive soil mapping; third, we evaluate the effectiveness of the proposed method in the designed experiment using the leave-one-out cross-validation method. In total, 191 evaluation cases are included in the experimental case base and the test results show that 58.7% of the topographic covariates originally used in each evaluation case are correctly selected by the proposed method, which suggests that the proposed most-similar-case method perform reasonably well even with a relatively limited size of the case base. The future work should include the selection of other types of environmental covariates (e.g., climate, organism, etc.) and the development of an automatic method to extract the existing application cases from literature. [ABSTRACT FROM AUTHOR]

Traditional machine learning models have mainly been used to study geological logging data of a single sample point, ignoring the fact that logging data has a strong spatial correlation. In this study, we use convolutional neural network to extract single-point features, structural features, and multidimensional features from logging data and compare the identification effects of lithology identification models based on the three features. The identification model based on the multidimensional feature extraction achieves 77.94% correctness in the test set, which is the best result among the identification models based on CNN and the three machine learning models. Based on this feature extraction model, the feature fusion modules in U-net and feature pyramid are added respectively to build two feature fusion models to combine the features extracted from different convolutional layers and improve the effectiveness of the model. The model also introduces attention mechanism to improve the role of useful features in the model training process. The identification accuracy of the two feature fusion models, U-CNN and P-CNN, reached 79.67% and 80.02% on the test set, respectively, which verified the effectiveness of the feature fusion models for lithology identification in the study area. [ABSTRACT FROM AUTHOR]