Integrating AI in legal analysis of satellite imagery : a focused approach using transformer models to guide classification

Tools

Rapach, Seonaid and Riccardi, Annalisa and Wheate, Rhonda (2025) Integrating AI in legal analysis of satellite imagery : a focused approach using transformer models to guide classification. In: Living Planet Symposium 2025, 2025-06-23 - 2025-06-27.

[thumbnail of Rapach-etal-LPS-2025-Integrating-AI-in-Legal-Analysis-of-Satellite-Imagery]
Preview
 Text. Filename: Rapach-etal-LPS-2025-Integrating-AI-in-Legal-Analysis-of-Satellite-Imagery.pdf
Accepted Author Manuscript
License: Creative Commons Attribution 4.0 logo
Download (2MB)| Preview

Abstract

Earth Observation (EO) imagery in an invaluable data source in a variety of humanitarian applications. Its adoption has been demonstrated in international courts, including the International Criminal Court (ICC) and the International Court of Justice (ICJ). In these proceedings, the most widely adopted method for analysing satellite imagery is visual interpretation, which typically involves object detection, such as identifying roadblocks​ [1]​, and classification, such as distinguishing and counting internally displaced people (IDP) settlement ​[2]​ and bombarded buildings ​[3]​. However, these methodologies are conducted solely with manual classification with interpreters​ [4]​, which is costly and time-consuming. With the rapid awareness and development of artificial intelligence (AI), every sector is motivated to assess possible applications, and the law is no different. In legal services, AI has already demonstrated its value in tasks such as analysing and drafting legal documents​ [5]​. However, discussions about the use of AI in judicial decision-making and law enforcement remain a subject of intense debate​ [6]​. Beyond these applications, there is also a clear opportunity for AI to aid in the processing and analysis of satellite imagery, along with other forms of spatial and imagery data. Integrating machine learning (ML) and deep learning (DL) models within legal methodological structures could improve the methodology, from reducing potential human errors to reducing the processing time, which can also ultimately allow for temporally and spatially scaling up the analysis. Currently, AI is used in a wide range of humanitarian applications, but there is justifiable concern over the prevalence of misclassifications from AI models. In international criminal courts, where admissible evidence must meet exacting standards, even modest false-positive rates can overestimate the presence or nature of key features, potentially rendering evidence inadmissible or unconvincing. Furthermore, limitations and concerns about AI need to be articulated and examined, including problems arising over the ‘black box’ data nature of AI decision-making, which must be thoroughly examined to address the specific needs of legal systems. Therefore, this research aims to outline the potential scope for integrating AI models for processing and analysing satellite imagery. More specifically, the research will outline how legal practitioners in international courts could adopt ‘transformer’ models to improve scene/object classification methods, without compromising the integrity of the evidence. Transformer models are particularly important for this application because ‘interventions’ can be applied to various, ensuring a robust but adaptable system​ [7]​. This approach allows practitioners to apply a single pre-pretrained model, but tailor it to suit their needs. Emphasis is placed on minimising either the rate of false positives or false negatives through targeted interventions, ensuring reliable and credible classifications. We will demonstrate this process through a case study where we will use a pre-trained transformer model for pixel-classification of forest disturbances​ [8]​, as if posed to detect instances of illegal deforestation. This scenario parallels the analytical challenges faced in human rights litigation, such as identifying instances of environmental destruction or land use violations from satellite imagery. Adopting a pre-trained model enables practitioners to utilise well-develop resources without requiring previous extra resources to train the model themselves, whilst also ensuring that this framework is transparent and reproducible. The following methodology adapts the Li ​[9]​ process, who performed targeted interventions to the most influential heads of a large-language model (LLM), to improve the ‘truthfulness’ of the model​ [9]​. Instead, we adopt this to the object classification of disturbed forests in Copernicus Sentinel-2 satellite imagery, using the visual transformer model developed by Schiller ​[8]​. This process enables evidence providers to adopt pre-trained models, whilst steering the behaviour of the model to minimise the instances of false-positive classifications of forest disturbances. False positive classification can undermine the integrity of the evidence, potentially rendering the litigation as false and destructive. Therefore, this process is essential for extracting expansive information on illegal deforestation, but also ensuring the that the evidence is robust and sound. In addition to the technical aspects, we will outline the implications of integrating this methodology from a legal and social perspective, including the rules of admissibility, ethics and accountability. This process provides a framework for integrating AI into legal workflows, offering a scalable and defensible approach to evidence analysis while meeting high evidentiary standards. ​[1] ​Prosecutor v. William Samoei Ruto and Jushua Arap Sang (ICC Transcript) ICC-01/09-01/11-T-109-ENG (9 April 2014). ​[2] ​Sufi and Elmi v. the United Kingdom (Judgment) no. 8319/07 and 11449/07 (ECtHR 28 November 2011). ​[3] ​Prosecutor v. Bosco Ntaganda (ICC Transcript) ICC-01/04-02/06-T-176-Red2-ENG (12 December 2016). ​[4] ​J. A. Quinn, M. M. Nyham, C. Navarro, D. Coluccia, L. Bromley and M. Luengo-Oroz, “Humanitarian applications of machine learning with remote-sensing data: review and case study in refugee settlement mapping,” Phil. Trans. R. Soc., vol. 376, no. 20170363, 2018. ​[5] ​I. Atrey, “Revolutionising the Legal Industry: The Intersection of Artificial Intelligence and Law,” International Journal of Law Management & Humanities, vol. 6, no. 3, pp. 1075-1089, 2023. ​[6] ​I. Taylor, “Justice by Algorithm: The Limits of AI in Criminal Sentencing,” Criminal Justice Ethics, vol. 42, no. 3, pp. 193-213, 2023. ​[7] ​J. Beal, E. Kim, E. Tzeng, D. H. Park, A. Zhai and D. Kislyuk, “Vision transformers in domain adaptation and domain generalization: a study of robustness,” 2020. [Online]. ​[8] ​C. Schiller, J. Költzow, S. Schwarz, F. Schiefer and F. E. Fassnacht, “Forest disturbance detection in Central Europe using transformers and Sentinel-2 time series,” Remote Sensing of Environment, vol. 315, no. 114475, 2024. ​[9] ​K. Li, O. Patel, F. Viegas, H. Pfister and M. Wattenberg, Inference-Time Intervention: Eliciting Truthful Answers from a Language Model, 2024. ​

ORCID iDs

Rapach, Seonaid, Riccardi, Annalisa ORCID logoORCID: https://orcid.org/0000-0001-5305-9450 and Wheate, Rhonda ORCID logoORCID: https://orcid.org/0000-0003-1604-5951;
CORE (COnnecting REpositories)