[论文解读] Greenhouse Segmentation on High-Resolution Optical Satellite Imagery using Deep Learning Techniques

Orkhan Baghirli, Imran Ibrahimli+1 more|arXiv (Cornell University)|Jul 22, 2020

Remote Sensing and LiDAR Applications参考文献 23被引用 4

一句话总结

本文提出了一种基于U-Net的深度学习框架，用于在阿塞拜疆1.5米分辨率的光学卫星影像中实现高精度的温室分割。通过整合空洞卷积、跳跃连接、双线性插值以及加权二元交叉熵 + Dice损失，最佳模型的F1分数达到93.29%，相比基线U-Net提升了4.48%。

ABSTRACT

Greenhouse segmentation has pivotal importance for climate-smart agricultural land-use planning. Deep learning-based approaches provide state-of-the-art performance in natural image segmentation. However, semantic segmentation on high-resolution optical satellite imagery is a challenging task because of the complex environment. In this paper, a sound methodology is proposed for pixel-wise classification on images acquired by the Azersky (SPOT-7) optical satellite. In particular, customized variations of U-Net-like architectures are employed to identify greenhouses. Two models are proposed which uniquely incorporate dilated convolutions and skip connections, and the results are compared to that of the baseline U-Net model. The dataset used consists of pan-sharpened orthorectified Azersky images (red, green, blue,and near infrared channels) with 1.5-meter resolution and annotation masks, collected from 15 regions in Azerbaijan where the greenhouses are densely congested. The images cover the cumulative area of 1008 $km^2$ and annotation masks contain 47559 polygons in total. The $F_1, Kappa, AUC$, and $IOU$ scores are used for performance evaluation. It is observed that the use of the deconvolutional layers alone throughout the expansive path does not yield satisfactory results; therefore, they are either replaced or coupled with bilinear interpolation. All models benefit from the hard example mining (HEM) strategy. It is also reported that the best accuracy of $93.29\%$ ($F_1\,score$) is recorded when the weighted binary cross-entropy loss is coupled with the dice loss. Experimental results showed that both of the proposed models outperformed the baseline U-Net architecture such that the best model proposed scored $4.48\%$ higher in comparison to the baseline architecture.

研究动机与目标

• 开发一种鲁棒的语义分割流程，用于识别高分辨率光学卫星影像中的温室。
• 解决温室检测中尺度变化、不规则形状、阴影和大气失真等挑战。
• 通过引入空洞卷积和双线性上采样等架构改进，提升基线U-Net的分割精度。
• 评估损失函数、数据增强和困难样本挖掘对模型泛化能力和性能的影响。
• 构建一种计算可行且可部署的解决方案，用于利用卫星数据进行气候智能型农业监测。

提出的方法

• 使用分辨率为1.5米的阿塞拜疆（SPOT-7）全色锐化、正射校正的卫星影像，包含RGB和近红外波段。
• 通过将图像分割为64×64像素的图像块（重叠32像素）并仅保留正类比例≥0.1的图像块，以解决类别不平衡问题。
• 通过计算NDVI和纹理通道增强特征，并应用数据增强技术，包括Mixup、旋转、翻转以及亮度/伽马调整。
• 设计了两种新型U-Net变体：Model A和Model B，通过引入空洞卷积和跳跃连接以捕捉多尺度上下文信息。
• 在解码器路径中用双线性插值替代或组合转置卷积，以减少棋盘纹伪影。
• 通过困难样本挖掘（HEM）、自适应学习率调度以及结合加权二元交叉熵和Dice损失的混合损失函数优化训练过程。

实验结果

研究问题

• RQ1 空洞卷积和跳跃连接在高分辨率卫星影像中如何提升温室分割性能？
• RQ2 用双线性插值替代转置卷积对分割质量与伪影减少有何影响？
• RQ3 加权二元交叉熵与Dice损失的组合相比标准损失函数，对F1分数和AUC分数有何影响？
• RQ4 数据增强和困难样本挖掘在阴影和雾霾等恶劣条件下在多大程度上提升了模型的鲁棒性？
• RQ5 与基线U-Net相比，架构上的差异在处理不规则、小型或被遮挡的温室时对性能有何影响？

主要发现

• 表现最佳的模型（Model B）取得了93.29%的F1分数，相比基线U-Net提升了4.48%。
• Model B在复杂场景中表现更优，例如被阴影覆盖或形状不规则的温室，其分割结果无间隙，能正确识别所有目标。
• 使用双线性插值替代转置卷积显著减少了棋盘纹伪影，并提升了分割掩膜的视觉质量。
• 混合损失函数（加权二元交叉熵 + Dice损失）实现了最高的AUC（95.91%）和F1分数，表明其具有出色的泛化能力和鲁棒性。
• 困难样本挖掘和数据增强显著提升了模型在恶劣条件下的性能，尤其在大气条件较差时表现更优。
• Model A表现居中，优于基线模型但逊于Model B，尤其在检测小型或部分被遮挡的温室时表现较弱。