Abstract
Land-cover classification in high-resolution remote sensing (HRS) images is a critical task for urban planning, environmental monitoring, and disaster management. However, existing deep learning models often face a trade-off between high computational costs and classification accuracy. This paper introduces AutoLC, a hybrid Neural Architecture Search (NAS) framework designed to automate the design of lightweight yet high-performing models. By combining a gradient-based searched encoder with an adaptive lightweight decoder, AutoLC achieves state-of-the-art results on the LoveDA benchmark while significantly reducing parameter counts and computational overhead compared to traditional manual architectures.
Introduction
Assigning a semantic label to every pixel in a satellite image—known as land-cover classification—is a cornerstone of remote sensing. While manual Convolutional Neural Networks (CNNs) like DeepLabV3+ and PSPNet have set high benchmarks for accuracy, they were originally designed for natural images (e.g., street scenes) and are often computationally redundant when applied to massive HRS datasets.
Researchers typically face an "efficiency vs. accuracy" dilemma. Manual models are accurate but heavy, while existing NAS methods often focus solely on the encoder, pairing it with rigid decoders that fail to recover the fine spatial details necessary for complex geographical scenes. AutoLC addresses this by searching for an efficient encoder while providing a flexible, adaptive decoder tailored to the multi-scale challenges of remote sensing.
Methodology
The AutoLC framework consists of two primary components designed to optimize both feature extraction and spatial reconstruction.
1. Automatically Searched Lightweight Encoder
Using a gradient-based search strategy similar to DARTS, the framework explores a two-level hierarchical search space:
- Micro-level (Cell): The system searches for the optimal combination of operations, such as convolutions, pooling, and skip connections, within a single cell.
- Macro-level (Architecture): The system determines the best pathway between different spatial resolutions to optimize the flow of information.
2. Adaptive Lightweight Decoder
Unlike standard decoders that use a fixed structure, AutoLC employs a design inspired by Feature Pyramid Networks (FPN). This adaptive decoder selects the most relevant feature maps from the searched encoder. It utilizes a Feature Fusion and Semantic Aggregation approach, merging information across the pyramid levels. To capture both local detail and global context, the model incorporates an Atrous Spatial Pyramid Pooling (ASPP) module at the final stage.
Experimental Setup and Datasets
The model was rigorously evaluated on the LoveDA dataset, a challenging benchmark containing 5,987 high-resolution images from three Chinese cities: Wuhan, Nanjing, and Changzhou. LoveDA is particularly demanding because it includes both urban and rural scenes, which exhibit vastly different spatial scales and spectral characteristics, providing a robust test for the model's generalization capabilities.
Key Results and Performance
AutoLC demonstrates that high-tier performance does not require massive model sizes. On the LoveDA validation set, AutoLC achieved a mean Intersection over Union (mIoU) of 45.26%, outperforming both manual benchmarks and previous NAS-based models.
Efficiency Comparison
When compared to the widely used PSPNet, AutoLC achieved superior accuracy with significantly fewer resources:
- Parameters: 4.28M vs. 46.59M (an 11x reduction).
- FLOPs: 102.2G vs. 709.4G (a 7x reduction).
Even the most compact variant of the model ($F=10, dim=64$) outperformed most manual benchmarks while utilizing only 3.68M parameters, proving the effectiveness of the search space.
Discussion and Insights
The search process revealed several surprising architectural preferences for remote sensing data:
- The Role of Pooling: In the optimal searched cell, more than half of the operations were pooling or skip connections, with convolutions making up only one-third. This suggests that efficient spatial aggregation is more critical for remote sensing than simply stacking deep convolutional layers.
- Resolution Preference: The algorithm consistently favored higher resolutions toward the end of the encoder. This "Stacked Hourglass" style helps minimize the resolution gap before upsampling, which is vital for preserving the boundaries of small objects like buildings and narrow roads.
- The "Goldilocks Zone": The authors identified a specific balance for model capacity. Increasing the number of channels or decoder width beyond a certain point leads to diminishing returns; if the ratio between these hyperparameters is unbalanced, the model becomes harder to converge, and performance drops.
Conclusion and Future Work
AutoLC provides a scalable and efficient solution for land-cover classification, bridging the gap between automated architecture design and the specific needs of remote sensing. While the search process remains computationally intensive (requiring approximately 8 days on a TITAN X GPU), the resulting models offer a significant leap in deployment efficiency.
Future research will focus on extending AutoLC to other dense prediction tasks, such as change detection and instance segmentation, and exploring how these searched architectures generalize across different sensor types, including multispectral and Synthetic Aperture Radar (SAR) imagery.
Summary updated