Knowledge Distillation-Based Lightweight Change Detection in High-Resolution Remote Sensing Imagery for On-Board Processing
Deep learning (DL) has been introduced to change detection (CD) due to its powerful feature representation and robust generalization abilities. However, the application of large DL models has high computational complexity and massive storage requirements for achieving good performance. For disaster...
| Main Authors: | , , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2024-01-01
|
| Series: | IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10403982/ |
| _version_ | 1797330422616555520 |
|---|---|
| author | Guoqing Wang Ning Zhang Jue Wang Wenchao Liu Yizhuang Xie He Chen |
| author_facet | Guoqing Wang Ning Zhang Jue Wang Wenchao Liu Yizhuang Xie He Chen |
| author_sort | Guoqing Wang |
| collection | DOAJ |
| description | Deep learning (DL) has been introduced to change detection (CD) due to its powerful feature representation and robust generalization abilities. However, the application of large DL models has high computational complexity and massive storage requirements for achieving good performance. For disaster emergency response and other applications with high timeliness requirements, it is difficult to deploy large DL models on spaceborne edge devices with limited resources to achieve on-board CD processing. To address this limitation, a novel CD based on knowledge distillation (CDKD) method that combines prototypical contrastive distillation and channel-spatial-normalized (CSN) distillation is proposed. PC distillation represents the feature distribution by calculating the differences between the similarities of pixel features and their positive and negative prototypes, and improves the student model's detection ability in changed regions that have similar features to the background by mimicking the relative feature distribution. CSN distillation combines two distillation paradigms, channel normalization and spatial normalization, and guides the student model to comprehensively learn the knowledge contained in the output probabilities of the teacher model to accurately identify changed regions with complex shapes. The effectiveness and reliability of the proposed CDKD method are verified on three public remote sensing CD datasets, and extensive experiments and analyses show that the proposed CDKD method can be used to train lightweight models with comparable performance to that of large models. |
| first_indexed | 2024-03-08T07:18:43Z |
| format | Article |
| id | doaj.art-af32fff66a85429d97d24367b8fcb5f6 |
| institution | Directory Open Access Journal |
| issn | 2151-1535 |
| language | English |
| last_indexed | 2024-03-08T07:18:43Z |
| publishDate | 2024-01-01 |
| publisher | IEEE |
| record_format | Article |
| series | IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing |
| spelling | doaj.art-af32fff66a85429d97d24367b8fcb5f62024-02-03T00:02:11ZengIEEEIEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing2151-15352024-01-01173860387710.1109/JSTARS.2024.335494410403982Knowledge Distillation-Based Lightweight Change Detection in High-Resolution Remote Sensing Imagery for On-Board ProcessingGuoqing Wang0https://orcid.org/0000-0002-0671-6777Ning Zhang1https://orcid.org/0000-0003-4717-2304Jue Wang2https://orcid.org/0000-0002-0229-8424Wenchao Liu3https://orcid.org/0000-0001-7747-523XYizhuang Xie4https://orcid.org/0000-0003-1312-2691He Chen5https://orcid.org/0000-0003-4182-6493National Key Laboratory of Science and Technology on Space-Born Intelligent Information Processing, Beijing Institute of Technology, Beijing, ChinaNational Key Laboratory of Science and Technology on Space-Born Intelligent Information Processing, Beijing Institute of Technology, Beijing, ChinaState Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan, ChinaNational Key Laboratory of Science and Technology on Space-Born Intelligent Information Processing, Beijing Institute of Technology, Beijing, ChinaNational Key Laboratory of Science and Technology on Space-Born Intelligent Information Processing, Beijing Institute of Technology, Beijing, ChinaNational Key Laboratory of Science and Technology on Space-Born Intelligent Information Processing, Beijing Institute of Technology, Beijing, ChinaDeep learning (DL) has been introduced to change detection (CD) due to its powerful feature representation and robust generalization abilities. However, the application of large DL models has high computational complexity and massive storage requirements for achieving good performance. For disaster emergency response and other applications with high timeliness requirements, it is difficult to deploy large DL models on spaceborne edge devices with limited resources to achieve on-board CD processing. To address this limitation, a novel CD based on knowledge distillation (CDKD) method that combines prototypical contrastive distillation and channel-spatial-normalized (CSN) distillation is proposed. PC distillation represents the feature distribution by calculating the differences between the similarities of pixel features and their positive and negative prototypes, and improves the student model's detection ability in changed regions that have similar features to the background by mimicking the relative feature distribution. CSN distillation combines two distillation paradigms, channel normalization and spatial normalization, and guides the student model to comprehensively learn the knowledge contained in the output probabilities of the teacher model to accurately identify changed regions with complex shapes. The effectiveness and reliability of the proposed CDKD method are verified on three public remote sensing CD datasets, and extensive experiments and analyses show that the proposed CDKD method can be used to train lightweight models with comparable performance to that of large models.https://ieeexplore.ieee.org/document/10403982/Change detection (CD)feature distributionknowledge distillation (KD)model compression and accelerationprobability distribution |
| spellingShingle | Guoqing Wang Ning Zhang Jue Wang Wenchao Liu Yizhuang Xie He Chen Knowledge Distillation-Based Lightweight Change Detection in High-Resolution Remote Sensing Imagery for On-Board Processing IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing Change detection (CD) feature distribution knowledge distillation (KD) model compression and acceleration probability distribution |
| title | Knowledge Distillation-Based Lightweight Change Detection in High-Resolution Remote Sensing Imagery for On-Board Processing |
| title_full | Knowledge Distillation-Based Lightweight Change Detection in High-Resolution Remote Sensing Imagery for On-Board Processing |
| title_fullStr | Knowledge Distillation-Based Lightweight Change Detection in High-Resolution Remote Sensing Imagery for On-Board Processing |
| title_full_unstemmed | Knowledge Distillation-Based Lightweight Change Detection in High-Resolution Remote Sensing Imagery for On-Board Processing |
| title_short | Knowledge Distillation-Based Lightweight Change Detection in High-Resolution Remote Sensing Imagery for On-Board Processing |
| title_sort | knowledge distillation based lightweight change detection in high resolution remote sensing imagery for on board processing |
| topic | Change detection (CD) feature distribution knowledge distillation (KD) model compression and acceleration probability distribution |
| url | https://ieeexplore.ieee.org/document/10403982/ |
| work_keys_str_mv | AT guoqingwang knowledgedistillationbasedlightweightchangedetectioninhighresolutionremotesensingimageryforonboardprocessing AT ningzhang knowledgedistillationbasedlightweightchangedetectioninhighresolutionremotesensingimageryforonboardprocessing AT juewang knowledgedistillationbasedlightweightchangedetectioninhighresolutionremotesensingimageryforonboardprocessing AT wenchaoliu knowledgedistillationbasedlightweightchangedetectioninhighresolutionremotesensingimageryforonboardprocessing AT yizhuangxie knowledgedistillationbasedlightweightchangedetectioninhighresolutionremotesensingimageryforonboardprocessing AT hechen knowledgedistillationbasedlightweightchangedetectioninhighresolutionremotesensingimageryforonboardprocessing |