Landcover Segmentation And Object Detection In Remote Sensing Images

Remote sensing technology is widely used in meteorological forecasting,urban planning,environmental monitoring,agricultural production assistance and so on.For massive remote sensing image data,the need for fast and effective intelligent analysis emerges as the times require.In view of the current application requirements of remote sensing technology in environmental monitoring,this paper conducts research from three aspects,which include land use segmentation,change detection and rotating target detection.The specific contents include:1.A large-scale remote sensing dataset of land use and land cover—MCPRL-LULC is constructed,which surpasses the existing public datasets in the rationality of category design and the extensiveness of applications and can be closely integrated with environmental monitoring.In addition,a benchmark land use segmentation algorithm is proposed based on this dataset,which is compared with multiple representative semantic segmentation networks,and the experimental results verify the effectiveness of the algorithm.2.In the task of change detection,a simple and effective change detection module based on self-attention mechanism is designed by using the inherent relationship between the two tasks of semantic segmentation and change detection,which can be easily inserted into the existing semantic segmentation network and transformed into a change detection network.Meanwhile,a data augmentation method based on serialized features is proposed,which can be embedded in any Transformer-based change detection network without modifying the labels.Experiments show that the proposed method has a greater performance advantage than other complex-designed change detection models.3.In the task of rotation object detection,a new one-stage anchor-free rotating object detection network based on FCOS and a new centerness calculation method are proposed,which can "soften" the centerness value,and make its distribution in the whole target area more reasonable.Meanwhile,a multi-center multi-stage rotating bounding box regression algorithm is proposed,and the multi-stage regression strategy from point to center and then to corner points is carried out in the regression branch.The experimental results on the public large-scale remote sensing dataset DOTA-v1.0 verify the effectiveness of the proposed method.