Using Contour Loss Constraint in Satellite Image Interpretation of Construction Disturbance

• Published in: IGARSS 2022 - 2022 IEEE International Geoscience and Remote Sensing Symposium pp. 2975 - 2978
• Main Authors: Lv, Ning, Yuan, Hao, Chen, Chen, Su, Tao, Deng, Jiaxuan, Zhou, Yang
• Format: Conference Proceeding
• Language: English
• Published: IEEE 17.07.2022
• Subjects: Buildings; Feature extraction; Local government; loss function; Mathematical models; parameter optimization; remote sensing interpretation; Satellites; semantic segmentation; Soil; Water conservation
• ISSN: 2153-7003
• DOI: 10.1109/IGARSS46834.2022.9883516

Accurate annotation of construction disturbance outline greatly influences the evaluation quality in interpretation application. While the loss functions used in semantic segmentation nowadays are difficult to reflect the contour information. Therefore, this paper proposes a RA U-Net(Residual Attention U-Net) construction disturbance interpretation. The model uses an exactly defined Contour Loss(CL) as the loss function. This function could reflect the local segmentation information of each element by extracting and transferring contour features into the weight matrix. Moreover, this paper modifies the CL function to make it suitable for construction disturbance interpretation. The experiments on M-nih Massachusetts Building Dataset and Construction Disturbance Dataset labeled manually show the capability of the proposed function in remote sensing interpretation. The proposed model has been used for the Soil and Water Conservation project supervised by the local government. And the IoU increased 1.85% maximum than binary cross-entropy(BCE) on Mnih Massachusetts Building Dataset.