Resumen de Evaluación de la eficiencia de predicción a deslizamientos usando un modelo cartográfico-hidrológico: Caso de estudio Cuenca la Carbonera, flanco SE del Volcán Pico de Orizaba.
Gabriel Legorreta Paulín, Rocío M. Alanís Anaya, Lilia Arana Salinas, José Fernanado Aceves Quesado
- español
En México hay una carencia de inventarios de deslizamientos y de su modelado. Como resultado tampoco existe una evaluación de las ventajas y la desventaja de los modelos de susceptibilidad de deslizamiento. En este trabajo se lleva a cabo una evaluación exhaustiva de la eficiencia de los modelos de susceptibilidad a deslizamientos utilizando la tecnología SIG. Para ello se emplean diferentes técnicas de comparación (característica operativa del receptor, eficiencia del modelo, razón de momios y distintas presiones del modelo obtenidas de una matriz de confusión) bajo el sistema LOGISNET y el programa estadístico SPSS. El área de estudio es la cuenca La Carbonera en el flanco SE del volcán Pico de Orizaba. En ella se realiza un inventario detallado de deslizamientos. Este inventario es el marco para la comparación cualitativa y cuantitativa con dos mapas de susceptibilidad creados con un modelo cartográfico-hidrológico (SINMAP). El modelo se ha seleccionado por ser considero en la literatura que su análisis tiene un éxito razonable en la definición de áreas que intuitivamente parecen ser susceptibles a deslizamientos en regiones con escasa información. Los resultados de este trabajo permiten establecer las bondades y limitaciones de las técnicas de comparación así como el de su uso en la evaluación de la eficiencia del modelo para predecir deslizamientos. La aplicación de estas técnicas de comparación permite observar la eficiencia de predicción del modelo cartográfico-hidrológico usando datos geotécnicos específicos del área de estudio y pre-establecidos por el sistema.
- English
Worldwide, the areas prone to gravitational processes (landslides) have been identified through the application of landslide susceptibility models incorporated into a Geographic Information System (GIS). These models use a variety of deterministic, heuristic, probabilistic and statistical methods at local or regional levels. In all cases, the evaluation of the efficiency of the model to predict the gravitational processes is key to ensure their reliability in the land-use management and planning. To this end, the susceptibility model should be compared with a detailed and accurate inventory of landslides. This research used the La Carbonera basin fluvial system, located in the southeast slope of the Pico de Orizaba volcano, Mexico, to evaluate the models and explain the comparison techniques. The study area has physiographic conditions that are susceptible to gravitational processes due to the combination of several factors, such as high rainfall during the rainy season, the intrinsic susceptibility of the existing types of rocks and deposits (poorly consolidated volcaniclastic deposits and weathered rocks, as well as folded and fractured sedimentary rocks), steep slopes and changes in land use. Within the study area, landslides and debris flows occur frequently. These two phenomena adversely affect human settlements and economic activities. In Mexico, there are few detailed maps on an inventory of landslides, geospatial databases of these processes, and susceptibility maps that could be used systematically to compare and contrast the limitations and advantages of the models for any region of the country. In addition, the evaluation of a model under natural conditions is a difficult issue due to the complexity of the natural and technical problems associated with it. The model performance may be compromised by natural issues such as the masking landslides by vegetation, the combination of several different types of landslides and processes, etc., in addition to technical issues including a poor DEM resolution, the sampling strategy selected, the collection and calculation of geotechnical parameters, the lack of a complete and detailed inventory, etc. The lack of a systematic comparison of the susceptibility models not only compromises the reliability of these models, but also leads to the abuse in their use. In order to alleviate some of these issues, this study built a detailed inventory of landslides based on two series of digital aerial photographs and field work. Landslides were digitized and incorporated into the GIS, and a geospatial database was developed for a better description. Field work and interpretation allowed the cartographic representation of 236 landslide processes. This inventory sets the framework for qualitative and quantitative comparisons, with two susceptibility maps created with a cartographic-hydrological model (SINMAP). SINMAP was selected because in the literature it is considered that its analysis is fairly successful in defining areas that intuitively seem to be susceptible to landslides in regions with scarce information.With the landslide inventory and the cartographic-hydrological susceptibility model, this work provides a comprehensive assessment of different comparison techniques to highlight the advantages and limitations of each. To this end, the model was run using specific geotechnical data for the study area and predefined geotechnical data. Different comparison techniques (receiver operational feature, overall model accuracy, odds ratio and different precision values obtained from a confounding matrix) were used under the LOGISNET system and the SPSS statistical program. The results of this work set the advantages and limitations of the comparison techniques, as well as their use in the assessment of the model efficiency to predict landslides. The assessment showed that the overall precision and the odds ratio should be evaluated in parallel with other precisions (user and producer precision, and model efficiency), as their values are strongly influenced by the most common class, usually those pixels that represent "stable areas". The characteristic receiver operating curve (ROC) is a better technique for use in an overall assessment of models, and can be used even to define precise and objective cut-off levels to classify maps into "stable" and "unstable" areas. The ROC curve and the producer precision showed a greater predictive capacity for SINMAP using geotechnical data specific to the study area (70.8% and 67.26%, respectively) relative to SINMAP using predefined geotechnical data (34.4% and 38.46%, respectively). The low user precision value for both, i.e. SINMAP using specific geotechnical data and SINMAP using predefined geotechnical data, was further confirmed by its large negative value in the model efficiency test in areas with landslides, showing that models poorly predict landslide pixels. This is due to the fact that the model tends to be over-predictive. The fieldwork supports that the susceptibility map drawn based on SINMAP with specific geotechnical data is considerably better than the one from SINMAP model with geotechnical data predefined by the system. The areas with over-prediction can potentially be affected by future landslides, as these areas are located on or near existing landslides.
