PaperlayerView exact revision on arXiv

Source-linked AI summary

A novel Deep Structure U-Net for Sea-Land Segmentation in Remote Sensing Images

Pourya Shamsolmoali, Masoumeh Zareapoor, Ruili Wang, Huiyu Zhou, Jie Yang

arXiv:2003.07784v1eess.IVcs.CV

TL;DR

Crowded remote-sensing scenes make sea–land segmentation difficult, and existing CNN performance is described as below the desired target. The paper proposes RDU-Net, which combines U-Net with densely connected residual blocks and short-range and identity-mapping connections. Experiments report that RDU-Net outperforms other state-of-the-art approaches on sea–land segmentation tasks.

  • Problem

    Sea–land segmentation remains challenging in complex remote-sensing scenes, and existing CNN performance is described as below the expected target.

    Source: §I · p. 1

  • Method

    RDU-Net extends U-Net with densely connected residual blocks and short-range and identity-mapping connections in its down- and up-sampling paths.

    Source: §I · p. 2

  • Results

    RDU-Net outperformed other state-of-the-art approaches on sea–land segmentation tasks.

    Source: §IV · p. 13

  • Takeaways & Limitations

    The authors report RDU-Net's performance, computational cost, and flexibility for sea–land segmentation tasks.

    Source: §IV · p. 13

Abstract

from arXiv · show

Sea-land segmentation is an important process for many key applications in remote sensing. Proper operative sea-land segmentation for remote sensing images remains a challenging issue due to complex and diverse transition between sea and lands. Although several Convolutional Neural Networks (CNNs) have been developed for sea-land segmentation, the performance of these CNNs is far from the expected target. This paper presents a novel deep neural network structure for pixel-wise sea-land segmentation, a Residual Dense U-Net (RDU-Net), in complex and high-density remote sensing images. RDU-Net is a combination of both down-sampling and up-sampling paths to achieve satisfactory results. In each down- and up-sampling path, in addition to the convolution layers, several densely connected residual network blocks are proposed to systematically aggregate multi-scale contextual information. Each dense network block contains multilevel convolution layers, short-range connections and an identity mapping connection which facilitates features re-use in the network and makes full use of the hierarchical features from the original images. These proposed blocks have a certain number of connections that are designed with shorter distance backpropagation between the layers and can significantly improve segmentation results whilst minimizing computational costs. We have performed extensive experiments on two real datasets Google Earth and ISPRS and compare the proposed RDUNet against several variations of Dense Networks. The experimental results show that RDUNet outperforms the other state-of-the-art approaches on the sea-land segmentation tasks.

I. INTRODUCTION

Sea–land segmentation remains difficult in crowded, high-resolution remote-sensing scenes, where complex intensity and texture distributions and objects such as ships or islands can disrupt classification. The paper introduces RDU-Net, combining densely connected residual blocks with U-Net to improve segmentation while limiting computational costs.

I. INTRODUCTION

  • Crowded, high-resolution scenes challenge sea–land segmentation because compound intensity and texture distributions, inland water, ships, and islands can confuse classifiers.The introduction notes that earlier approaches increased accuracy but were not stable under these conditions.Source: §I · p. 1
  • DeepUNet can perform poorly on smooth sea–land boundaries or complex structures, motivating a new end-to-end, pixel-wise architecture.The paper names that architecture Residual Dense U-Net (RDU-Net), using convolution layers and multi-scale densely connected residual blocks.Source: §I · p. 2
  • RDU-Net combines residual-network feature reuse with DenseNet feature detection, using short-range and identity-mapping connections across feature layers.The authors say the ordered connections are intended to shorten backpropagation distance and reduce computation costs.Source: §I · p. 2
  • The proposed dense blocks have L log L runtime complexity, compared with O(L2) for standard DenseNets at depth L.The paper also states that this design does not increase GPU memory relative to ordinary DenseNets.Source: §I · p. 2
  • RDU-Net combines U-Net with densely connected residual blocks and is reported to improve results while remaining more efficient than U-Net, FusionNet, and DeepUNet.The authors report comparisons among these networks and describe RDU-Net as far more efficient than the other state-of-the-art networks.Source: §I · p. 2

II. RELATED WORK

Related work spans classical feature-based and thresholding approaches, learning-based segmentation, and deep encoder–decoder networks. The paper positions its method among these approaches by adding densely connected residual blocks to both U-Net sampling paths.

II. RELATED WORK

  • Remote-sensing segmentation work also covers classical machine learning, clustering and ensembles, multilayer models, and convolutional approaches.The cited examples include random forests, hierarchical segmentation, ensemble clustering, multilayer perceptrons, and feature-pyramid networks.Source: §II · pp. 2–3
  • Prior sea–land methods include thresholding, hand-engineered texture and intensity features, level sets, and supervised learning, while composite imagery can still lead to misclassified pixels.Examples include green or shadowed land classified as one region and water noise or waves treated as land.Source: §II · p. 3
  • Deep learning motivates pixel-wise segmentation, but FCNs can produce coarse maps after subsampling; residual learning and encoder–decoder models are among the approaches discussed.The section names EddyNet as an encoder–decoder model for pixel-wise classification.Source: §II · p. 3
  • U-Net links down- and up-sampling paths using feature maps from the contracting path, while DeepUnet concatenates contracting-path layers to form an expansive path.The proposed model adds DenseNet blocks with multilevel convolution layers and combined connections in both sampling paths.Source: §II · p. 3

III. PROPOSED METHOD

This section introduces the detailed structure of the proposed model.

III. PROPOSED METHOD

  • The section presents the detailed structure of the proposed model.The supplied passage provides no further method details.Source: §III · p. 3

A. Overall Architecture

RDU-Net builds on U-Net’s contracting and expansive paths, adding dense residual blocks and identity connections within and across paths. Its design also specifies how features are downsampled, upsampled, and merged across levels.

A. Overall Architecture

  • RDU-Net extends U-Net’s contracting and expansive paths with proposed DenseNet blocks and identity links, including long links between the paths.Each dense block also has several identity-mapping connections within its path.Source: A. Overall Architecture · p. 3, Figure 1 · p. 4, Figure 2 · p. 5
  • RDU-Net’s convolution design uses full preactivation, combining convolution layers with batch normalization and PReLU.The paper states that Adamax is used to derive the weights and biases.Source: A. Overall Architecture · p. 4
  • Each expansive level upsamples features and merges them with the corresponding contracting-path features through an identity-mapping connection.Downsampling uses stride-2 convolutions instead of max-pooling, while upsampling layers replace U-Net’s deconvolution blocks.Source: A. Overall Architecture · p. 4
  • A residual unit adds its learned residual function to an identity-mapped input, and deeper features accumulate earlier residual functions.The paper uses PReLU after the element-wise addition.Source: A. Overall Architecture · pp. 4–5
  • Convolution layers on both sides of each dense block connect feature maps to the block and help keep the network balanced.The channel count increases after contracting-path downsampling and is split across expansive levels.Source: A. Overall Architecture · p. 5
  • To enlarge receptive fields systematically, RDU-Net adopts a cited model; the paper notes that very large receptive fields can introduce noise.The discussion also identifies stacking layers and subsampling as ways to enlarge receptive fields.Source: A. Overall Architecture · p. 5

B. Densely Connected Residual Network

The proposed block combines dense feature exploration with residual feature reuse, using shorter backpropagation paths within RDU-Net. Its components are reported to improve segmentation results and robustness over U-Net, DeepUNet, and FusionNet.

B. Densely Connected Residual Network

  • The double path combines dense connections that explore new features with residual connections that reuse shared features before a final transformation.The paper describes this data flow as aggregating and passing features toward mapping or prediction.Source: B. Densely Connected Residual Network · pp. 5–6
  • Each RDU-Net dense block has six convolution layers, short-range connections, identity mapping, and batch normalization after the final convolution.The first and last kernels are 1×1; the remaining kernels are 3×3.Source: B. Densely Connected Residual Network · p. 6
  • The growth rate increases with convolution depth, adding more feature maps from later layers without introducing additional hyperparameters.The paper gives the growth rate as p = 2^(m−1)p_0.Source: B. Densely Connected Residual Network · p. 6
  • The proposed DenseNet uses O(L log L) connections rather than DenseNet’s O(L^2), with maximum backpropagation distance 1 + log_2 L.Each layer receives inputs from at most log(i) + 1 earlier layers.Source: B. Densely Connected Residual Network · p. 7
  • Short-range connections and identity mappings merge low- and high-level features, improving RDU-Net’s segmentation results and robustness over U-Net, DeepUNet, and FusionNet.The combined feature maps are then fed to the upscaling layers.Source: B. Densely Connected Residual Network · p. 7

IV. EXPERIMENTS DETAILS AND ANALYSIS

The section introduces the datasets, experimental results, and comparisons with other models.

IV. EXPERIMENTS DETAILS AND ANALYSIS

  • The experiments section covers the datasets, experimental results, and comparisons between the proposed model and other models.Source: §IV · p. 7

A. Data augmentation and preprocessing

The study uses manually cropped Google Earth sea–land images alongside ISPRS_Benchmark data, then divides the images into train, validation, and test sets. Data augmentation addresses insufficient training samples.

A. Data augmentation and preprocessing

B. Network Setup and Experiments

The authors compare RDU-Net with U-Net, FusionNet, and DeepUNet using the same datasets and experimental settings. They report the implementation environment and use Softmax for pixel classification.

B. Network Setup and Experiments

C. Evaluation Metrics

F1 score and global pixel-wise accuracy evaluate sea-land segmentation, with F1 summarizing precision and recall. The reported metrics are derived from pixel-based confusion matrices.

C. Evaluation Metrics

  • Training uses 300 epochs with mini-batches of 16, an initial learning rate of 0.001 reduced tenfold every 15 epochs, and Adamax optimization.Source: C. Evaluation Metrics · p. 8
  • F1 score is the harmonic mean of precision and recall, alongside global pixel-wise accuracy as an evaluation metric.Overall accuracy is calculated by normalizing the confusion matrix trace.Source: C. Evaluation Metrics · p. 8
  • The metrics are calculated from pixel-based confusion matrices, and overall accuracy is the normalized trace of the confusion matrix.Source: C. Evaluation Metrics · p. 8

D. Performance and Comparison

RDU-Net is evaluated against segmentation models on the whole dataset and on complex sea-land scenes. The reported results show strong segmentation performance, with analyses of its dense blocks, identity mapping, and computational efficiency.

D. Performance and Comparison

  • On the Fig. 5 island scene, RDU-Net segments sea and land without errors and exceeds U-Net, FusionNet, and DeepUNet on precision, accuracy, recall, and F1.Source: D. Performance and Comparison · p. 8, Figure 5 · p. 9, Table III
  • On the more complicated Fig. 6 scene, RDU-Net’s accuracy is 6.51% higher than U-Net, 5.44% higher than FusionNet, and 2.76% higher than DeepUNet.Its F1 is 2.73% higher than DeepUNet, 4.77% higher than FusionNet, and 5.87% higher than U-Net.Source: D. Performance and Comparison · p. 9, Figure 6 · p. 9, Table IV · p. 10
  • The authors attribute RDU-Net’s stronger performance to densely connected residual blocks with shorter backpropagation distances and identity mappings that support feature reuse.Source: D. Performance and Comparison · p. 9
  • In Figs. 7 and 8, RDU-Net segments tiny ships and minor objects; U-Net and FusionNet perform poorly, while DeepUNet misses some boundaries and shadow areas.The test scenes include close-set small ships, shadows, green spaces, and tiny sea-land boundaries.Source: D. Performance and Comparison · p. 10, Figure 7 · p. 10, Figure 8 · p. 10
  • With the proposed DenseNet blocks, RDU-Net improves performance and reaches its highest rate in fewer iterations, which the authors associate with reduced computation cost.Source: D. Performance and Comparison · p. 11, Figure 9 · p. 11
  • K is highest for dense blocks 1–4 in the downsampling path and block 9 in the upsampling path, and lowest for downsampling blocks 5–8.The authors interpret the low values as indicating that the convolution layers’ dimension-increasing role is more important than the DenseNet blocks themselves in those blocks.Source: D. Performance and Comparison · p. 11, Figure 11 · p. 11
  • RDU-Net achieves 97.39% overall accuracy and 97.13% precision, the best performance among the compared deep neural network models on the whole dataset.The comparison is reported in Table VII.Source: Table VII · p. 11, D. Performance and Comparison · p. 12
  • RDU-Net requires approximately half the parameters and FLOPs of original DenseNet for comparable accuracy, and has better computation efficiency than DeepUNet and RU-Net.The comparison covers the Massachusetts roads and sea-land datasets.Source: D. Performance and Comparison · p. 13, Table VIII · p. 13

IV. CONCLUSION

RDU-Net combines standard U-Net with densely connected residual blocks and short-range and long identity-mapping connections for sea-land segmentation. The authors report that it outperformed other state-of-the-art approaches across multiple quality metrics.

IV. CONCLUSION

  • RDU-Net outperformed other state-of-the-art approaches across a number of quality metrics for sea-land segmentation.The reported comparisons include U-Net, FusionNet, DeepUnet, and other models.Source: §IV · p. 13
  • RDU-Net combines standard U-Net with densely connected residual blocks and short-range and long identity-mapping connections.The architecture is intended to extract deep features and hierarchically reuse them for end-to-end image segmentation.Source: §IV · p. 13
  • The authors plan to make RDU-Net more efficient, improve segmentation accuracy, and reduce computation costs in future work.They intend to extend the accuracy goal beyond sea-land segmentation to other segmentation tasks.Source: §IV · p. 13

Related papers

Loading 2003.07784v1…