Quantum-Enhanced Vision Transformer for Flood Detection using Remote Sensing Imagery

Soumyajit Maity and Behzad Ghanbarian · 2026

Read Quantum-Enhanced Vision Transformer for Flood Detection using Remote Sensing Imagery by Soumyajit Maity and Behzad Ghanbarian (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-13689.txt.

Introduction

Soumyajit Maity and Behzad Ghanbarian introduce a Quantum-Enhanced Vision Transformer for flood detection from remote sensing imagery. Classical deep learning models struggle with the high-dimensional, nonlinear complexity of such data. Their hybrid processes inputs through a ViT backbone and a quantum branch in parallel. The quantum branch uses a 4-qubit parameterized circuit for localized feature mapping. Fused representations feed a binary classifier for flood versus no flood. The hybrid raised accuracy from about 84 percent to about 94 percent.

Catalog record

CREATORSoumyajit Maity and Behzad Ghanbarian
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE21971
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2603.13689

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

Quantum-Enhanced Vision Transformer for Flood Detection using Remote Sensing Imagery Soumyajit Maity and Behzad Ghanbarian Abstract Reliable flood detection is critical for disaster management, yet classical deep learning models often struggle with the high-dimensional, nonlinear complexities inherent in remote sensing data. To mitigate these limitations, we introduced a novel Quantum-Enhanced Vision Transformer (ViT) that synergizes the global context-awareness of transformers with the expressive feature extraction capabilities of quantum computing. Using remote sensing imagery, we developed a hybrid architecture that processes inputs through parallel pathways, a ViT backbone and a quantum branch utilizing a 4-qubit parameterized quantum circuit for localized feature mapping. These distinct representations were fused to optimize binary classification. Results showed that the proposed hybrid model significantly outperformed a classical ViT baseline, increased overall accuracy from 84.48% to 94.47% and the F1-score from 0.841 to 0.944. Notably, the quantum integration substantially improved discriminative power in complex terrains for both class. These findings validate the potential of quantum-classical hybrid models to enhance precision in hydrological monitoring and earth observation applications. I Introduction Flooding is among the most devastating natural hazards worldwide, responsible for substantial loss of life and economic damage amounting to tens of billions of dollars annually. The increasing frequency and intensity of extreme precipitation events, driven in part by climate change and rapid urbanization, have further exacerbated flood risks across both developed and developing regions. Consequently, reliable and timely flood detection has become a critical component of disaster risk reduction, emergency response, and long-term infrastructure planning. Accurate flood detection enables authorities and stakeholders to issue early warnings, allocate resources effectively, and implement mitigation strategies that can significantly reduce human and economic losses. Despite its importance, flood detection remains a complex and challenging task. It requires high-quality hydrological and geospatial data describing the physical and environmental characteristics of a given area, including topography, land use and land cover, soil properties, drainage networks, and meteorological forcing. These variables interact in highly nonlinear ways across multiple spatial and temporal scales, making the underlying processes difficult to model using purely physics-based or empirical approaches. As a result, a wide range of data-driven and hybrid methodologies have been proposed to improve flood detection and mapping, particularly in data-rich environments [ 1 ] . In recent years, artificial intelligence (AI) techniques have emerged as powerful tools for flood detection, benefiting from the growing availability of remote sensing data, in situ observations, and high-performance computing resources. Classical machine learning (ML) models, such as support vector machines, random forests, and gradient boosting, have been applied to classify flooded and non-flooded regions using multi-source datasets [ 23 , 18 ] . More recently, deep learning (DL) architectures, including convolutional neural networks and recurrent neural networks, have demonstrated strong performance in extracting spatial and temporal patterns from satellite imagery [ 22 , 25 ] . These approaches have significantly advanced the state of the art by improving detection accuracy, automating feature extraction, and enabling near-real-time flood monitoring. Nevertheless, classical ML and DL methods face fundamental limitations when dealing with extremely high-dimensional feature spaces, complex nonlinear dependencies, and large-scale datasets. Issues such as computational cost, model generalization, and sensitivity to data imbalance and noise remain persistent challenges. Furthermore, the expressive power of classical models may be insufficient to capture subtle correlations in certain hydrological systems, particularly when the underlying dynamics are governed by multiscale interactions and rare extreme events. These challenges motivate the exploration of alternative computational paradigms that can potentially enhance learning efficiency and representational capacity. In this context, quantum machine learning (QML) has recently emerged as a promising research direction at the intersection of quantum computing and AI. QML aims to leverage quantum mechanical principles, such as superposition, entanglement, and quantum interference to enhance classical learning algorithms [ 5 ] . Within the QML framework, classical input features are mapped into quantum states using a quantum feature map, typically encoded into qubits or qudits [ 12 ] . Quantum circuits then act on these states, transforming them into richer representations in high-dimensional Hilbert spaces. Measurements of the resulting quantum states are used to compute observables, kernels, or cost functions, which are subsequently optimized using classical routines. This embedding of classical data into quantum Hilbert spaces constitutes a key advantage of QML, as it enables access to exponentially large feature spaces that are intractable for classical computers to represent explicitly [ 21 ] . As a result, QML models have the potential to uncover complex patterns and correlations that may be difficult to detect using conventional ML or DL techniques. From a theoretical perspective, QML offers opportunities for achieving quantum advantage in learning tasks, particularly in problems characterized by high dimensionality, strong nonlinearities, and limited training data. Several classes of QML approaches have been proposed in the literature. Quantum kernel methods, which are currently the most mature and compatible with noisy intermediate-scale quantum (NISQ) hardware [ 4 ] , use quantum computers to evaluate similarity measures between data points in quantum feature space. Variational quantum circuits (VQCs) rely on parameterized quantum circuits whose parameters are trained through classical optimization [ 2 ] , enabling flexible and expressive models analogous to neural networks. A third category, quantum-enhanced feature extraction, embeds quantum components within classical ML pipelines to improve feature representation while retaining classical training frameworks [ 6 ] . Despite rapid methodological developments in QML, its application to geoscience and hydrology, particularly flood detection, remains largely unexplored. To date, most QML studies have focused on benchmark datasets or synthetic problems, with only limited attention given to real-world environmental systems. This gap highlights a significant opportunity to investigate whether quantum-enhanced learning models can offer tangible benefits for flood detection, especially in scenarios involving high-dimensional remote sensing data and complex hydrological interactions. Fig. 1: The framework proposed in this study. II Methodology The proposed framework in this study was structured into four distinct phases, started with data acquisition followed by a rigorous pre-processing pipeline to standardize the inputs. Subsequently, we implemented a Quantum-enhanced ViT architecture to leverage hybrid feature extraction. In the final phase, the model was trained and its performance was compared to a classical ViT baseline to validate the efficacy of the quantum integration. Figure 1 provides an overview of the proposed experimental workflow. Both the proposed Quantum-enhanced ViT and the classical ViT base design use the same ViT backbone configuration, 1024 hidden dimensions, 24 transformer encoders, 16 focus heads, 4096 intermediate MLP size and patch size of 14\times 14 , which gives approximately 304 million trainable parameters for the ViT alone. Crucially, both models were initialized with no pre-trained weights, which ensures that any performance differences between the two architectures can only be attributed to the structural contribution of the quantum-measurement segment and the fusion classifier, and not to the transfer of the learning benefits of the large-scale training. Full final fine-tuning was carried out on all layers of the transformer encoder for both models. No layer freezing or partial fine-tuning schedule was used, which allowed the focus masses to adapt from random start to the spectral and spatial characteristics of the SAR and multispectral imaging. All computations, including training and inference, were conducted on a high-performance workstation equipped with an Intel(R) Xeon(R) w7-3555 processor, 192 GB of

Read the complete text