Special Edition: Advanced Computational Methods in Cultural Heritage Preservation
This Article is archived at
| Universidade Federal de Minas Gerais (UFMG), Belo Horizonte, Brasil |
Special Editor:
Dr. Ana Carolina Silva Santos, PhD
Departamento de Antropologia e Arqueologia
Universidade Federal de Minas Gerais (UFMG)
Original Research Article
Author:
Er. Rishabh Aryan
Department of Computer Science and Engineering
Indian Institute of Information Technology, Bhagalpur, Bihar, India
Email: rishabh.250201011@iiitbh.ac.in
Abstract
The Brazilian Amazon harbours several thousand documented geoglyphs, earthworks, and rock-art sites that represent some of the most complex yet least-understood pre-Columbian cultural productions in the Americas. Manual documentation is impeded by the scale of the territory, forest canopy occlusion, and the scarcity of trained archaeologists who also hold the linguistic and cultural competence required for contextualised interpretation. This paper presents the design, training, and evaluation of a Neuro-Symbolic Multimodal Learning (NSML) framework that integrates convolutional feature extraction, Vision Transformer (ViT-L/16) encoders, and an indigenous-knowledge ontology to classify and partially interpret Amazonian geoglyphs and rock art. The model is trained on a curated dataset of 14,372 georeferenced image instances drawn from satellite synthetic aperture radar (SAR) imagery, aerial LiDAR point clouds, and ground-level RGB photography, spanning 47 verified archaeological sites across the states of Amazonas, Pará, Acre, and Mato Grosso. On the held-out test partition, NSML achieves an F1-score of 88.2% and an AUC of 0.95, outperforming five baseline deep-learning architectures. Gradient-weighted Class Activation Mapping (Grad-CAM) and attention rollout are used to generate spatially localised explanations. A governance framework grounded in the principles of Free, Prior, and Informed Consent (FPIC) and the CARE Principles for Indigenous Data Governance is formalised, addressing data sovereignty, differential access tiers, and benefit-sharing mechanisms. Our findings demonstrate that neuro-symbolic approaches not only surpass purely connectionist models in classification accuracy but also produce explanations that are more readily validated by indigenous knowledge holders, thereby strengthening community trust and cultural sovereignty over their own heritage data.
Keywords: Neuro-Symbolic AI, Geoglyphs, Amazonian Rock Art, Vision Transformer, Explainability, Indigenous Data Sovereignty, FPIC, CARE Principles, LiDAR, Remote Sensing
DOI: 10.37066/ralap.2026.ufmg.spec.01
Citation:
Er. R. Aryan, “Neuro-Symbolic Deep Learning for Interpreting Pre-Columbian Geoglyphs and Rock Art in the Brazilian Amazon: Implementation of a Multimodal Model, Explainability Analysis, and Governance Frameworks for Indigenous Cultural Sovereignty,” Revista Latinoamericana de la Papa, vol. 1, no. Special Edition: Advanced Computational Methods in Cultural Heritage Preservation, pp. 1–9, 2026. DOI: 10.37066/ralap.2026.ufmg.spec.01
Received: 17/05/2026
Accepted: 30/05/2026
Published: 01/06/2026
Copyright of this article is retained by the author.