Source-linked AI summary
Quantum-Inspired Convolutional Networks for High-Resolution Pseudo-Label Generation from Multimodal Earth Observation Data
Abhishek Singh, Giovanni Cuozzo, Alexander Jacob, Claudia Notarnicola
TL;DR
High-resolution land-cover mapping is constrained by scarce detailed annotations, while multimodal Earth-observation imagery requires robust representations under coarse supervision. The paper proposes a quantum-inspired encoder–decoder with reconstruction and auxiliary classification to generate pseudo-labels from Sentinel-1 and Sentinel-2 imagery guided by coarse MODIS information. On DFC2020, the model is reported to perform competitively and often outperform reference methods despite using fewer parameters.
Problem
High-resolution land-cover mapping is constrained by limited detailed annotations, and multimodal EO data require robust spectral–spatial representations under weak or coarse supervision.
Source: §1 · p. 1Method
The paper proposes a parameter-efficient quantum convolution–based encoder–decoder with reconstruction and an auxiliary classifier for high-resolution pseudo-label generation.
Source: §1 · p. 2Results
The proposed model performs competitively and often outperforms reference methods on DFC2020 despite using fewer parameters.
Source: §4 · p. 5Takeaways & Limitations
The lightweight, parameter-efficient architecture shows potential for compact onboard AI in real-time Earth-observation applications.
Source: §4 · p. 5
Abstract
from arXiv · showhide
Accurate land-cover mapping from multimodal Earth observation (EO) data remains a challenge in regions where high-resolution annotations are scarce or costly to obtain. This work introduces a quantum convolutional encoder-decoder network for pseudo-label generation using Sentinel-1 SAR and Sentinel-2 multispectral imagery, guided only by coarse MODIS-derived land-cover maps. The proposed framework uses quantum-inspired convolutional blocks to enhance feature extraction and cross-sensor fusion, enabling the model to infer a fine-scale semantic structure from heterogeneous inputs. Experiments were conducted on the DFC2020 dataset, which provides a quadruple of Sentinel-1 SAR, Sentinel-2 multispectral, low-resolution MODIS labels and high-resolution reference land-cover maps. The proposed training framework first learns a robust representation of multimodal inputs through the reconstruction branch before any supervision from low-resolution labels is applied. This self-supervised stage enables the model to capture a fine-scale spectral-spatial structure, which is then used to generate high-resolution pseudo-labels. These pseudo-labels are subsequently evaluated through classification metrics on the test data, demonstrating that the proposed approach produces more spatially coherent and class-consistent compared to classical baselines. The results highlight the potential of quantum-inspired deep learning to improve label generation in low-annotation remote sensing scenarios, offering a scalable path towards high-quality land-cover products for EO applications.
1 Introduction
High-resolution land-cover mapping is important for environmental monitoring, climate analysis, and land-use planning, but detailed annotations are scarce and multimodal inputs require robust learning under coarse supervision. The paper proposes a parameter-efficient quantum-inspired encoder–decoder and a reconstruction module with an auxiliary classifier for high-resolution pseudo-label generation.
1 Introduction
- High-resolution land-cover mapping supports environmental monitoring, climate analysis, and sustainable land-use planning, but detailed annotated datasets are limited.Source: §1 · p. 1
- Multimodal Sentinel-1 SAR and Sentinel-2 imagery provide complementary information, yet effective use requires robust spectral–spatial representations under weak or coarse supervision.Source: §1 · p. 1
- Conventional CNNs may require substantial labeled data, use large parameter budgets, and struggle to generalize across remote-sensing distributions, sensors, and acquisition conditions [1].Source: §1 · p. 1
- The proposed approach combines a parameter-efficient quantum convolution–based encoder–decoder with a reconstruction module and auxiliary classifier to generate high-resolution pseudo-labels.The reconstruction module learns spectral–spatial representations and uses reconstructed features for pseudo-label generation.Source: §1 · p. 2
- The paper compares its approach with state-of-the-art methods in segmentation accuracy.Source: §1 · p. 2
2 Proposed Methodology
The methodology embeds locally pooled image features in a six-qubit parameterized circuit and places quantum-enhanced blocks across a hybrid encoder–decoder. Reconstruction and auxiliary classification are used to learn representations and support masked pseudo-label generation and segmentation.
2 Proposed Methodology
- Each QuantumConv layer pools non-overlapping 2 × 2 patches by channel-wise mean, adjusts the descriptor to six values, and angle-encodes them into a six-qubit circuit.Descriptors with fewer than six channels are zero-padded; those with more are truncated to their first six components.Source: §2 · p. 2
- The quantum circuit embeds features through single-qubit RY rotations, applies parameterized entangling layers, and produces nonlinear features through Pauli-Z measurements.The circuit uses a ring entanglement pattern and has depth L = 4.Source: §2 · p. 2
- The hybrid encoder–decoder applies classical convolutions and seven QBlocks: three in the encoder, one at the bottleneck, and three in the decoder.Each QBlock contains two QuantumConv layers; quantum outputs are projected into classical feature space before later convolutional processing.Source: §2 · pp. 2–3
- Training uses reconstruction and auxiliary classification to learn representations, generate masked pseudo-labels, and support semantic segmentation.The reconstruction branch preserves fine-grained spatial and spectral structure, while auxiliary classification learns semantic information from coarse labels and produces temperature-scaled confidence maps.Source: §2 · p. 3
3 Experimental Results and Discussions
Experiments on DFC2020 assess the proposed approach using five remapped land-cover classes and compare it with reference methods. The proposed model achieves the best overall segmentation performance among the compared architectures while using a more compact model, with quantum-circuit simulation adding training overhead.
3.1 Experimental Data and Model Setup and Loss Function Formulation
- Experiments used 4,615 training, 513 validation, and 986 test patches, evaluating five dominant remapped classes from DFC2020.The original 17 MODIS-derived classes were remapped to a unified 10-class scheme; the selected classes were Forest, Grassland, Croplands, Urban, and Water.Source: §3.1 · p. 3
- Training combines a hybrid L1-SSIM reconstruction loss, auxiliary classification loss, and a downstream segmentation loss combining label-smoothed cross-entropy with Dice loss.The reconstruction loss balances pixel-level and perceptual fidelity; the segmentation loss targets pixel accuracy, region overlap, and class imbalance.Source: §3.1 · pp. 3–4
- PQC parameters retain nonzero gradients during training, although gradient magnitudes decrease in deeper quantum blocks, indicating a vanishing-gradient tendency.The parameters remain trainable and are updated through standard backpropagation.Source: §3.1 · p. 4
3.2 Qualitative and Quantitative Evaluation and Discussions
- The proposed method achieves the best overall segmentation performance among the compared architectures while using a substantially more compact model.Quantum-circuit simulation adds training overhead, so the paper frames its practical advantage as parameter efficiency and accuracy rather than raw speed.Source: Table 2 · p. 4, §3.2 · p. 5
- Self-supervised reconstruction pre-training and auxiliary classification support masked high-resolution pseudo-label generation from image patches.The reconstruction branch learns from high-resolution inputs, while the auxiliary branch regularizes latent representations for semantic consistency across reconstructed and masked regions.Source: Figure 3 · p. 3, §3.2 · p. 4, Figure 4 · p. 4
- UNet and TransUNet reference methods were trained using the same data partitions, input patch size, batch size, and loss function.The methods were adapted to the dataset's input dimensionality and number of output classes.Source: §3.2 · p. 4
4 Conclusion
The paper presents a quantum-integrated convolutional encoder-decoder for generating high-resolution pseudo-labels from Sentinel-1 and Sentinel-2 imagery using coarse MODIS land-cover information. On DFC2020, it reports competitive and often better results than reference methods despite using fewer parameters, and identifies compact onboard deployment as a potential application.
4 Conclusion
- On DFC2020, the proposed model is reported as competitive and often better than reference methods despite using fewer parameters.The paper also reports significant improvement in accuracy metrics without providing specific values in this passage.Source: §4 · p. 5
- The lightweight, parameter-efficient design is presented as having potential for compact onboard AI deployment in real-time EO applications.Source: §4 · p. 5