针对目前遥感影像中烤烟边缘识别效率低且识别精度低等问题，文章提出一种融合轻量化ASPP和U-Net框架的遥感影像烤烟种植区域提取模型。首先，该模型在U-Net编码层和解码层连接处加入轻量化空洞空间金字塔池化模块;其次，该模型将线性整流函数(Rectified Linear Unit，ReLU)替换为ReLU6 激活函数，能够在低精度计算时压缩动态范围，从而使算法更具鲁棒性;最后，该模型通过采用形态学孔洞填充构建标签图后处理算法，实现分割结果优化。为验证该模型框架的有效性和适用性，文章采用无人机遥感影像作为实验数据集，构建与传统语义分割模型的对比实验以及消融实验等。实验结果表明，通过与FCN、U-Net、SegNet和DeepLabV3+等传统语义分割算法相比较，文章提出的模型获得了较好的分割效果，其像素准确率和平均交并比分别为 93。7%和 84。1%。此外，该模型在保证模型精度的情况下，还能够提高模型的计算速度。
Flue Cured Tobacco Area Extraction of Remote Sensing Image by Integrating Lightweight ASPP and U-Net
To address the issues of low efficiency and accuracy for flue cured tobacco area extraction in remote sensing images,we propose a novel tobacco area extraction model with remote sensing images.The proposed model integrates the lightweight ASPP and U-Net framework for improving performance.The model enhances its performance in three ways:1)integrating a lightweight atrous spatial pyramid pooling module at the junction of the U-Net encoding and decoding layers,2)substituting the ReLU activation function with ReLU6,which can compress the dynamic range during low-precision computation and enhance the algorithm's robustness,and 3)optimizing segmentation results by developing a post-processing algorithm that constructs labeled maps through morphological hole filling.To demonstrate the effectiveness of the model,we utilize UAV remote sensing images as the experimental dataset.We also conduct comparative experiments with traditional semantic segmentation models and ablation study of the proposed model.Compared to traditional semantic segmentation algorithms such as FCN,U-Net,SegNet and DeepLabV3+,the proposed method achieves better performance,where the pixel accuracy and the mean intersection over union of the proposed model are 93.7%and 84.1%,respectively.Furthermore,the model enhances computational speed while maintaining the accuracy of segmentation.
tobacco planting area extractionlightweight ASPPU-Netpost-processing
万方数据
维普
