Uncertainty Gated Network for Land Cover Segmentation

Pascual, Guillem; Seguí, Santi; Vitrià, Jordi

doi:10.1109/cvprw.2018.00052

Cited by 9 publications

(14 citation statements)

References 11 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…A classical neural network is proposed in [17] that optimizes the Jaccard index for the land cover classification problem. An Uncertainty Gated Network [47] is proposed that models the multi-scale contexts by leveraging the heteroscedastic measure of uncertainty for the classification of all pixels of a satellite image. The Dense Fusion Classmate Network (DFCNet) [48] incorporates mid-level information by using an auxiliary road dataset in addition to the deepglobe dataset [49] for land cover classification.…”

Section: Related Workmentioning

confidence: 99%

Deep Hybrid Network for Land Cover Semantic Segmentation in High-Spatial Resolution Satellite Images

2021

View full text Add to dashboard Cite

Land cover semantic segmentation in high-spatial resolution satellite images plays a vital role in efficient management of land resources, smart agriculture, yield estimation and urban planning. With the recent advancement in remote sensing technologies, such as satellites, drones, UAVs, and airborne vehicles, a large number of high-resolution satellite images are readily available. However, these high-resolution satellite images are complex due to increased spatial resolution and data disruption caused by different factors involved in the acquisition process. Due to these challenges, an efficient land-cover semantic segmentation model is difficult to design and develop. In this paper, we develop a hybrid deep learning model that combines the benefits of two deep models, i.e., DenseNet and U-Net. This is carried out to obtain a pixel-wise classification of land cover. The contraction path of U-Net is replaced with DenseNet to extract features of multiple scales, while long-range connections of U-Net concatenate encoder and decoder paths are used to preserve low-level features. We evaluate the proposed hybrid network on a challenging, publicly available benchmark dataset. From the experimental results, we demonstrate that the proposed hybrid network exhibits a state-of-the-art performance and beats other existing models by a considerable margin.

show abstract

Section: Related Workmentioning

confidence: 99%

Deep Hybrid Network for Land Cover Semantic Segmentation in High-Spatial Resolution Satellite Images

2021

View full text Add to dashboard Cite

show abstract

“…4 showing the end-to-end pipeline of the DDCM-Net combined with a pre-trained model for land cover classification. Compared to other encoder-decoder architectures, our proposed DDCM-Net only fuses low-level features one time before the final prediction CNN layers, instead of aggregating multi-scale features captured at many different encoder layers [2], [6], [27], [31], [32], [33], [34], [35], [36]. This makes our model simple and neat, yet effective with lower computational cost.…”

Section: The Ddcm Networkmentioning

confidence: 99%

“…In our work, we only utilize the first three bottleneck layers of pretrained ResNet-based [42] backbones (both ResNet50 [42] and SE-ResNeXt50 [43]) and remove the last bottleneck layer and the fully connected layers to reduce the number of parameters to train. Furthermore, due to the larger complexity and variety of the DeepGlobe dataset compared to the ISPRS data, we utilize a DDCM(s = 2) module configured with larger dilation growing rates [1,2,4,8,16,32] as the low-level encoder, and two DDCM(g = 2, s = 2) modules configured by [1,2,4] and [1] as the high-level decoder. This configuration results in feature maps of size 64-channel and 32-channel, rather than 36-channel and 18-channel for the model on ISPRS data.…”

Section: The Ddcm Networkmentioning

confidence: 99%

“…We used two different dilation settings strategies. For ISPRS data, we configured a DDCM module with linearly growing dilation rates ( [1,2,3,5,7,9]) as the low-level features encoder, while for the DeepGlobe dataset, we build the encoder with exponentially growing dilation rates ( [1,2,4,8,16,32]) with a stride of two, since we see that the DeepGlobe images contain more spatially chaotic objects with lower resolutions, larger scales and less geometrical attributes than the ISPRS images. We believe DDCM modules with bigger dilation configurations could capture larger multi-scale and global context in this case, but require more computational cost.…”

Section: B Flexible Dilation and Density Policiesmentioning

confidence: 99%

See 1 more Smart Citation

Dense Dilated Convolutions’ Merging Network for Land Cover Classification

Liu

Kampffmeyer

Jenssen

et al. 2020

IEEE Trans. Geosci. Remote Sensing

114

View full text Add to dashboard Cite

Land cover classification of remote sensing images is a challenging task due to limited amounts of annotated data, highly imbalanced classes, frequent incorrect pixel-level annotations, and an inherent complexity in the semantic segmentation task. In this work, we propose a novel architecture called the Dense Dilated Convolutions Merging Network (DDCM-Net) to address this task. The proposed DDCM-Net consists of dense dilated image convolutions merged with varying dilation rates. This effectively utilizes rich combinations of dilated convolutions that enlarge the network's receptive fields with less parameters and features compared to the state-of-the-art approaches in the remote sensing domain. Importantly, DDCM-Net obtains fused local and global context information, in effect incorporating surrounding discriminative capability for multi-scale and complex shaped objects with similar color and textures in very high resolution aerial imagery. We demonstrate the effectiveness, robustness and flexibility of the proposed DDCM-Net on the publicly available ISPRS Potsdam and Vaihingen data, as well as the DeepGlobe land cover dataset. Our single model, trained on 3-band Potsdam and Vaihingen data, achieves better accuracy in terms of both mean intersection over union (mIoU) and F1-score compared to other published models trained with more than 3band data. We further validate our model on the DeepGlobe dataset, achieving state-of-the-art result 56.2% mIoU with much less parameters and at a lower computational cost compared to related recent work.

show abstract

“…However, ultra-high resolution images cannot always fit in the limited GPU memory in practice. A popular solution to this problem is to first crop the original image into smaller patches for prediction [13,[20][21][22] and then merge them together. In GLNet [13], the cropped patches are used for extracting local features.…”

Section: Introductionmentioning

confidence: 99%

Learning Skip Map For Efficient Ultra-High Resolution Image Segmentation

Pi¹,

Jiang

Xiong

2021

2021 IEEE International Conference on Image Processing (ICIP)

View full text Add to dashboard Cite

The pattern distribution of ultra-high resolution images is usually unbalanced. While part of an image contains complex and fine-grained patterns such as boundaries, most areas are composed of simple and repeated patterns. In this work, we propose to learn a skip map, which can guide a segmentation network to skip simple patterns and hence reduce computational complexity. Specifically, the skip map highlights simple-pattern areas that can be down-sampled for processing at a lower resolution, while the remaining complex part is still segmented at the original resolution. Applied on the stateof-the-art ultra-high resolution image segmentation network GLNet, our proposed skip map saves more than 30% computation while maintaining comparable segmentation performance.

show abstract

Uncertainty Gated Network for Land Cover Segmentation

Cited by 9 publications

References 11 publications

Deep Hybrid Network for Land Cover Semantic Segmentation in High-Spatial Resolution Satellite Images

Deep Hybrid Network for Land Cover Semantic Segmentation in High-Spatial Resolution Satellite Images

Dense Dilated Convolutions’ Merging Network for Land Cover Classification

Learning Skip Map For Efficient Ultra-High Resolution Image Segmentation

Contact Info

Product

Resources

About