Integrated Deep Learning Based Segmentation And Classification Method For Boundary Delineation Of Agricultural Fields In Multitemporal Satellite Images
The multi-temporal satellite data gives the data in periodic basis which helps for continuous monitoring, but due to the earth’s rotation, climatic changes, sensor characteristics, etc. There are too many distortions and noises which have to be removed before further processing for getting better results. In this study, a multispectral segmentation method for automated delineation of agricultural field boundaries in remotely sensed images is presented. In most image semantic segmentation tasks, the Deep Learning has shown its irreplaceable advantages, specifically the U-Net network architecture. With this motivation, in this work proposed a learning-based, simultaneous segmentation and classification method based on the Deep Multi-scale U-Net (DMU-Net) structure with deformable convolutional layers. In addition using cellular automata based Gaussian filtering is used to enhance the multitemporal images, to remove "noise" and to enhance the major direction.The second approach is linear feature extraction based on some parameters such as higher STD, length, strength and contrast. In this way, information from several spectral bands can be used for delineating field borders with different characteristics. Finally the experimental results reveal that the proposed segmentation method is more efficient than the existing segmentation techniques such as EGKL method and field boundary segmentation method in factors of each quantitative overall performance metrics and appropriateness for land-use classification.