Scholarlink

Abstract <p>This study investigates model adaptation strategies using gradient boosting (LightGBM) for solving inverse problems in exploration geophysics focused on reconstructing spatial distributions of subsurface properties from surface measurements. Addressing limitations of traditional approaches requiring exhaustive site-specific training, we implement four model adaptation strategies: incremental training, pruning and extending, source-target model ensemble, and feature space augmentation. Experiments employ three synthetic datasets of progressive geological complexity—homogeneous (Easy), pattern variable (Medium), and spatially heterogeneous (Hard), representing a Norilsk-type four-layer section with gravimetric, magnetic, and magnetotelluric measurements. Results demonstrate that gradient boosting achieves solution accuracy comparable to neural networks while reducing computational costs. At the same time, the use of gradient boosting adaptation allows one to further reduce computational costs. The study creates an effective basis for geophysical research in conditions of limited data.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Gradient Boosting Ensemble Adaptation for Geophysical Inverse Problems under Data Scarcity

  • 期刊论文

  • 发表日期：

    2025年12月1日

查看全文
  • M. K. Shaleev,
  • I. V. Isaev,
  • I. E. Obornev,
  • S. A. Dolenko

摘要

Abstract

This study investigates model adaptation strategies using gradient boosting (LightGBM) for solving inverse problems in exploration geophysics focused on reconstructing spatial distributions of subsurface properties from surface measurements. Addressing limitations of traditional approaches requiring exhaustive site-specific training, we implement four model adaptation strategies: incremental training, pruning and extending, source-target model ensemble, and feature space augmentation. Experiments employ three synthetic datasets of progressive geological complexity—homogeneous (Easy), pattern variable (Medium), and spatially heterogeneous (Hard), representing a Norilsk-type four-layer section with gravimetric, magnetic, and magnetotelluric measurements. Results demonstrate that gradient boosting achieves solution accuracy comparable to neural networks while reducing computational costs. At the same time, the use of gradient boosting adaptation allows one to further reduce computational costs. The study creates an effective basis for geophysical research in conditions of limited data.