New AI Network Enhances UAV Thermal Mapping in Dynamic Environments

Researchers have introduced MAPS-Net, a multi-modal attention and planning network designed to improve autonomous navigation for unmanned aerial vehicles (UAVs) in dynamic environments. Published in the journal Cluster Computing in October 2026, the study presents a system that couples deep representation learning with Monte Carlo tree search to address the challenge of informative path planning for multi-UAV teams. The primary objective of the network is to facilitate the autonomous mapping of thermal hotspots, a task that requires complex decision-making under constraints such as tight flight budgets and limited communication capabilities among the vehicles.

The network integrates an egocentric belief grid, footprint-level sensor statistics, and mission context through a series of processing stages. These include gated preprocessing, multi-scale feature extraction, and multi-head self-attention mechanisms. This fusion results in two key outputs: a masked action policy and a bounded value estimate. These outputs respectively guide and evaluate a PUCT search, which operates on an analytic expected-information-gain model of the team’s belief state. Notably, the complete system is trained through self-play, eliminating the need for human demonstrations during the training phase.

The study evaluated the performance of MAPS-Net using experiments conducted on a real airborne thermal orthomosaic. The results indicated that the proposed network matches the performance of a strong greedy information-gain planner regarding residual map entropy. However, MAPS-Net demonstrates significant advantages over other baselines in terms of detection accuracy. Specifically, the network surpassed coverage and random baselines by 18.5 percent and 88 percent, respectively, in relative hotspot F1 score.

The researchers observed that the system scales superlinearly as the number of vehicles increases, with performance improvements noted when expanding the team from two to eight UAVs. Furthermore, the network exhibits resilience to communication disruptions. The study found that MAPS-Net loses under two entropy points when communication is removed entirely, indicating that the local planning capabilities remain effective even in the absence of inter-vehicle data sharing.

The evaluation also included tests under non-stationary field conditions, which introduce additional complexity through drifting temperatures, cloud shadows, and moving no-fly zones. In these dynamic scenarios, MAPS-Net delivered the best detection performance among all planners tested. Additionally, the network attained the best mapping and detection scores at fine planning resolutions.

By combining deep learning with search-based planning, MAPS-Net offers a viable solution for the efficient mapping of thermal hotspots.

Subject of Research: Autonomous UAV navigation and thermal mapping using multi-modal attention networks and Monte Carlo tree search

Article Title: MAPS-Net: a multi-modal attention and planning network with Monte Carlo tree search for autonomous UAV navigation in dynamic environments

Article References: Wang, Z., Luo, X., Wang, Z., & Liu, S. (2026). MAPS-Net: a multi-modal attention and planning network with Monte Carlo tree search for autonomous UAV navigation in dynamic environments. Cluster Computing, 29(16), Article 922. https://doi.org/10.1007/s10586-026-06688-4

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06688-4

Keywords: UAV navigation, Monte Carlo tree search, Thermal mapping, Multi-modal learning, Autonomous systems, Path planning, Deep learning, Cluster Computing, MAPS-Net, multi-modal, attention, planning