Improved semantic segmentation method using edge features for winter wheat spatial distribution extraction from Gaofen-2 images

Abstract

In the final feature map obtained using a convolutional neural network for remote sensing image segmentation, there are great differences between the feature values of the pixels near the edge of the block and those inside the block; ensuring consistency between these feature values is the key to improving the accuracy of segmentation results. The proposed model uses an edge feature branch and a semantic feature branch called the edge assistant feature network (EFNet). The EFNET model consists of one semantic branch, one edge branch, one shared decoder, and one classifier. The semantic branch extracts semantic features from remote sensing images, whereas the edge branch extracts edge features from remote sensing images and edge images. In addition, the two branches extract five-level features through five sets of feature extraction units. The shared decoder sets up five levels of shared decoding units, which are used to further integrate edge features and deep semantic features. This strategy can reduce the feature differences between the edge pixels and the inner pixels of the object, obtaining a per-pixel feature vector with high inter-class differentiation and intra-class consistency. Softmax is used as the classifier to generate the final segmentation result. We selected a representative winter wheat region in China (Feicheng City) as the study area and established a dataset for experiments. The comparison experiment included three original models and two models modified by adding edge features: SegNet, UNet, and ERFNet, and edge-UNet and edge-ERFNet, respectively. EFNet's recall (91.01%), intersection over union (81.39%), and F1-Score (91.68%) were superior to those of the other methods. The results clearly show that EFNET improves the accuracy of winter wheat extraction from remote sensing images. This is an important basis not only for crop monitoring, yield estimation, and disaster assessment but also for calculating land carrying capacity and analyzing the comprehensive production capacity of agricultural resources.

CSU-Net: Contour Semantic Segmentation Self-Enhancement for Human Head Detection

“…Li et al [20], developed a framework to improve the semantic segmentation results by decoupling features into the body and the edge parts to handle inner object consistency and finegrained boundaries jointly. Equally, Yin et al [21] proposed a model that links a branch of edge features and a branch of semantic features to ensure consistency between these feature values. This model clearly improves the precision of segmentation results.…”

Polarimetric Synthetic Aperture Radar Image Semantic Segmentation Network with Lovász-Softmax Loss Optimization

“…The fully convolutional neural network proposed by Long [27] achieves segmentation of images of any size by replacing the full connection layer of CNN with the convolution layer, which has already been developed and applied in SAR image segmentation [28,29]. Semantic segmentation attracts a significant amount of attention in the development of deep learning for achieving pixel-wise segmentation [30][31][32][33][34][35][36][37][38][39]. ENet [30], proposed by Paszke et al, has been improved and applied in remote sensing image semantic segmentation [31].…”

“…ERFNet [32], proposed by Romera et al, adopts None-bottleneck-1D to enhance the learning ability of the network and speed up the segmentation process. EFNet is proposed by Yin et al [33] based on ERFNet, and the two networks are compared for winter-wheat spatial distribution extraction from Gaofen-2 images. DeepLabV3 [34], proposed by Chen et al, applies modules employing atrous convolution in cascade or parallel to capture multi-scale context by adopting multiple atrous rates.…”