Explainable Remote Sensing Classification using XAI-Guided Large Language Models
13.59 MB
suljettu
Julkaisu on tekijänoikeussäännösten alainen. Teosta voi lukea ja tulostaa henkilökohtaista käyttöä varten. Käyttö kaupallisiin tarkoituksiin on kielletty.
Lataukset6
Pysyvä osoite
Verkkojulkaisu
DOI
Tiivistelmä
Deep learning (DL) models have proven very effective in land cover classifications in remote sensing (RS). However, the black-box nature of these models limits their interpretability and leads to lower trust in their results; this is especially true for parameters related to decision-making and the need for transparency in large-scale operations. Although there are many explainable artificial intelligence (XAI) techniques that visualize how an example from a model works, these methods often create output that is either too technical or too difficult for anyone other than a model expert to interpret. To overcome this shortcoming, this thesis develops a unified framework to integrate XAI techniques with large language models (LLMs) in order to create automatically generated natural-language explanations of DL model classifications for satellite images. The framework consists of two primary components: the first is the development of a methodology for translating visual outputs from XAI techniques into a structured text representation of these outputs; the second is to input this text into an LLM so that the LLM can generate a concise natural-language explanation that will link the visual features identified by each XAI and how those features affect the result of a classification. In addition, a combined XAI experiment is conducted, where the outputs of all three methods are merged into a single prompt to capture complementary spatial and feature-level information.
To evaluate explanation quality, an LLM-as-a-judge framework was adopted with multi-criteria scoring based on specificity and usefulness, combined with pairwise and per-class comparison analysis. Experimental results on the EuroSAT dataset demonstrate that the proposed framework significantly improves explanation quality, achieving a win rate of 75.83\% and a +18.50\% improvement over the baseline when the three XAI methods are combined. These findings highlight the importance of integrating complementary XAI signals to enhance the clarity, reliability, and usability of LLM-generated explanations for non-expert users in RS applications.