Research on the Optimization of Sample Point Placement for Ground Substrate Survey based on Interpretable Machine Learning

To obtain comprehensive and accurate data, it is necessary to conduct field surveys of ground substrate, arable land, soil quality, and other natural resource classes. This is in difficult and key areas. Manual sweeping can improve survey accuracy, but the investment in human and material resources, and the long survey cycle can complicate taking into account the impact of geological, hydrological, soil, and other factors on the deployment of sample points. In addition, technical experts can be relied on for expertise, professionalism, and a scientific approach. However, generalizability is weak due to its subjective nature. Machine learning can form an efficient and scientific scheme to deploy sample points. A large amount of data, such as soil, geology, and meteorology, can be simulated using machine learning models to generate effective multidimensional data features for engineering, identify potential patterns, and optimize sample point locations and quantities based on nonlinear and higher-order interactions. In contrast to traditional black box models that use prediction accuracy as a single metric, interpretable machine learning (IML) provides insight into how the model learns and makes predictions. As a result, we will be able to identify potential biases and errors more easily and build more reliable sample point placement models. In this study, we proposed the RF-GA algorithm, which integrates environmental factor data often relied on by experts’ judgments, such as topography, soil type, elevation, slope, surface cover, and historical soil texture. By using genetic algorithms to filter variables, the feature space can be optimized. A plan is developed for surface matrix survey points in northeast Tieling using intermediate scale factor modeling. There are a variety of ecosystems and topography types in the Tieling area, which lies between the Liaohe Plain and Liaodong mountainous hilly area. An optimized plan for regional ground substrate survey point placement was finally determined after nearly 110,000 calculations. The optimized scheme generated by the RF-GA algorithm and the expert manual scheme (1:250,00 survey accuracy) are similar by 87%, and the RF-GA rule adopts an iterative progression of prime points of a certain surface cover, and the point placement scheme reflects improved refinement and sampling point representativeness. As part of the exploration of similar environment judgments, the scheme optimized 785 points (swept survey) to 136, utilizing zonal classification and hierarchical sample point deployment. The RF-GA algorithm found that the main factors affecting point deployment are topography, historical soil texture, and slope. As both plain and hilly terrain are considered in the study area, RF-GA has high generalizability and can be applied to plains, hills, and transition zones between them. We have demonstrated the feasibility of RF-GA sample point placement based on interpretable machine learning, which is valuable for ground substrate surveys and point optimization research.