基于HRNet-OCR模型的农村宅基地提取方法

Rural homestead extraction from remote sensing images based on HRNet- OCR model

关键词:  农村宅基地  无人机影像  深度学习  HRNet-OCR  语义分割

Key words:  rural homestead  UAV image  deep learning  HRNet-OCR  semantic segmentation

摘要:

【目的】 农村宅基地信息统计是制定农村宅基地制度改革政策方向的基础，目前，基于遥感影像的农村宅基地提取还主要停留在人工目视解译的阶段，这种传统的提取方法效率低、成本高、耗时长，基于遥感影像自动化提取农村宅基地的相关研究较少。【方法】 文章收集了德清县无人机遥感影像数据，建立了训练集、验证集和测试集，构建HRNet-OCR模型，并与FCN、UNet、DeepLabV3Plus这3种模型在不同场景下进行对比。【结果】 模型精度评价指标IoU表明，在平原和丘陵地区HRNet-OCR比FCN、UNet和DeepLabV3Plus分别高了4.24%、3.72%和2.82%，在山区HRNet-OCR比FCN、UNet和DeepLabV3Plus分别高了3.59%、2.77%和1.55%，且模型在边缘细节上表现得更优秀。【结论】 基于HRNet-OCR识别模型使得遥感影像农村宅基地提取更为准确，具有更好的鲁棒性，可为精准提取农村宅基地提供重要参考价值。未来更快速、高效的高精度提取方法还有待进一步研究。

Abstract:

[Purpose] Rural Homestead information statistics is the basis for formulating the policy direction of rural homestead system reform. At present，rural homestead extraction based on remote sensing images mainly stays in the stage of manual visual interpretation. The traditional extraction method has low efficiency，high cost and long time-consuming. There are few studies on Automatic extraction of rural homestead based on remote sensing images.[Method] This paper collects the remote sensing image data of UAV in Deqing County，establishes the training set，verification set and test set，constructs the HRNet-OCR model，and compares it with FCN，UNET and deeplabv3plus in different scenarios.[Result] The model accuracy evaluation index IOU shows that in plain and hilly areas，HRNet-OCR is 4.24%，3.72%and 2.82%higher than FCN，UNET and deeplabv3plus respectively，and in mountainous areas，HRNet-OCR is 3. 59%，2.77%and 1.55%higher than FCN，UNET and deeplabv3plus respectively，and the model performs better in edge details.[Conclusion] Based on HRNet-OCR recognition model，rural homestead extraction from remote sensing images is more accurate and robust，which can provide important reference value for accurate extraction of rural homestead. In the future，more rapid，efficient and high-precision extraction methods need to be further studied.