Contributed Session 8: Methods and Models for Earth & Geohazards

Land cover mapping is vital for understanding land use, biodiversity, and ecosystem dynamics. While advancements in remote sensing and AI have improved mapping accuracy, challenges like data heterogeneity, high acquisition costs, and annotation labor remain. This study proposes a cost-effective methodology to develop a reproducible semantic segmentation dataset for agricultural landscapes based on: (1) acquiring and georeferencing high-resolution satellite imagery; (2) segmenting and validating regions with expert annotations for accurate class identification; and (3) training deep learning models, including U-Net, SegNet, and DeepLabV3, with and without pre-trained backbones, to evaluate their performance. DeepLabV3 consistently outperformed other models. The study also evaluates the impact of pre-trained backbones on DeepLabV3’s performance, with Receiver Operating Characteristic (ROC) analysis confirming its strong ability to differentiate land cover classes. The findings support sustainable land management and informed decision-making, demonstrating the transformative potential of deep learning and satellite imagery for agricultural applications.

Predicting seismic events is a complex challenge requiring the integration of multiple analytical methods to enhance forecasting accuracy. This study explores anomaly detection in geophysical data to improve earthquake prediction models. We employ a three-part approach: trend change analysis, wavelet analysis, and anomaly detection using Variational Autoencoders (VAE). Trend change analysis identifies shifts in geophysical patterns that may signal seismic activity, while wavelet analysis decomposes data into frequency components to detect localized anomalies over time. VAEs, a machine learning technique, help identify and quantify subtle anomalies that could serve as preseismic indicators.Our results suggest that certain anomalies can be detected days before a seismic event, providing insights into potential earthquake precursors. Integrating these anomalies into predictive models may improve forecasting accuracy, offering a complementary approach to traditional methods. However, further validation is needed to assess the consistency and reliability of these findings.

The Earth’s crust behaves as a complex, non-linear dynamic system that is characterized by long-range correlations. As an expression of this complex system, seismic phenomena also show long-range correlations, critical instabilities, multifractal hierarchical structures, for which standard equilibrium Boltzmann-Gibbs statistical mechanics is not applicable. In this work we apply the probability distribution of the seismic magnitude derived in the framework of the non-extensive statistical mechanics in different seismotectonic settings and we show that this distribution – called q-exponential – performs better than the classical exponential distribution obtained from the well-known Gutenberg-Richter law. Moreover variations of its q parameter can be associated with phase changes of the seismic cycle.