[论文解读] Satellite Imagery Feature Detection using Deep Convolutional Neural Network: A Kaggle Competition

Vladimir Iglovikov, Sergey Mushinskiy+1 more|arXiv (Cornell University)|Jun 19, 2017

Remote-Sensing Image Classification参考文献 8被引用 148

一句话总结

This paper adapts a multispectral U-Net-like Fully Convolutional Network for semantic segmentation of satellite imagery, introduces a joint loss with boundary handling, and achieves third place in the DSTL Kaggle competition without heavy ensembling.

ABSTRACT

This paper describes our approach to the DSTL Satellite Imagery Feature Detection challenge run by Kaggle. The primary goal of this challenge is accurate semantic segmentation of different classes in satellite imagery. Our approach is based on an adaptation of fully convolutional neural network for multispectral data processing. In addition, we defined several modifications to the training objective and overall training pipeline, e.g. boundary effect estimation, also we discuss usage of data augmentation strategies and reflectance indices. Our solution scored third place out of 419 entries. Its accuracy is comparable to the first two places, but unlike those solutions, it doesn't rely on complex ensembling techniques and thus can be easily scaled for deployment in production as a part of automatic feature labeling systems for satellite imagery analysis.

研究动机与目标

• Motivate accurate pixel-level classification of diverse satellite imagery classes for mapping, monitoring, and disaster response.
• Adapt fully convolutional networks to multispectral remote sensing data and evaluate data fusion strategies.
• Develop training objectives and pipeline adjustments to handle boundary effects and data imbalance.
• Assess the role of reflectance indices and augmentation in improving segmentation performance, especially for under-represented classes.

提出的方法

• Adaptation of a fully convolutional network (U-Net) to multispectral inputs with early fusion of RGB, M-band, and reflectance indices.
• Use a joint loss combining binary cross-entropy with a differentiable approximation of the Jaccard index (IoU) for training segmentation masks.
• Employ data augmentation including Dih4 transformations and patch-based training with 112x112 inputs and 128-channel batches.
• Address boundary effects by adding a cropping layer to output and reflections to padding areas to mitigate edge artifacts.
• Train separate models per class for the 8 target classes and use a patch-based prediction strategy to manage limited dataset size and class imbalance.
• Evaluate with the Jaccard index (IoU) on public and private test sets and report per-class results.

实验结果

研究问题

• RQ1 Can a multispectral U-Net with early fusion of multiple spectral bands achieve competitive semantic segmentation on small satellite datasets?
• RQ2 What training objective adjustments (e.g., IoU-inspired loss) improve pixel-wise segmentation quality in imbalanced, multispectral remote sensing data?
• RQ3 How do boundary effects from patch-based predictions impact accuracy, and can architectural tweaks mitigate them without heavy ensembling?
• RQ4 Do reflectance indices (e.g., CCCI, NDWI) complement learned features for under-represented classes such as waterways and standing water?
• RQ5 Is per-class separate modeling advantageous over a single multi-class model in this setting?

主要发现

• The approach achieved third place out of 419 entries on the Kaggle DSTL Satellite Imagery Feature Detection challenge.
• Joint loss combining binary cross-entropy and a differentiable IoU surrogate improved segmentation training.
• Separately modeling the first six classes and using border-cropping techniques reduced boundary artifacts and improved edge predictions.
• Reflectance indices helped segment water-related classes and vegetation, with indices performing better for under-represented classes like waterways.
• Per-class models with patch-based inputs yielded competitive results without heavy model ensembling.
• Per-class IoU scores varied substantially, with Waterway 0.9697 (Public) and 0.9131 (Private), Standing water 0.6081 (Public) and 0.5272 (Private), and some large vehicles showing very low scores (e.g., Vehicle Large: 0.2964 Public, 0.0331 Private).