Abdullah Azeem
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Abdullah Azeem has a total of 12 co-authors. Here are the top co-authors by number of shared publications:
- Zhengzhou Li
College of Microelectronics and Communication Engineering, Chongqing University, Chongqing, China
9 shared publications - Abubakar Siddique
School of Microelectronics and Communication Engineering, Chongqing University, Chongqing, China
8 shared publications - Yuting Zhang
School of Microelectronics and Communication Engineering, Chongqing University, Chongqing, China
8 shared publications Wenhao Chen
School of Microelectronics and Communication Engineering, Chongqing University, Chongqing, China
4 shared publications- Yongsong Li
School of Artificial Intelligence, Chongqing Technology and Business University, Chongqing, China
3 shared publications - Dong Cao
Computational Aerodynamics Institute, China Aerodynamics Research and Development Center, Mianyang 621000, China
2 shared publications
Latest Publications
- 10 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 63 (01962892) , 1-23
Prototype-Guided Multilayer Alignment Network for Few-Shot Object Detection in Remote Sensing
Few-shot object detection (FSOD) addresses the challenge of limited labeled data by enabling detectors to learn from minimal annotations. Recent work on image–text fusion has shown promise in overcoming data scarcity issues. However, these models suffer from catastrophic forgetting due to two distinct challenges during few-shot adaptation. First, feature space distortion disrupts established relationships between image–text modalities and blurs decision boundaries, leading to inaccuracies in object classification. Second, attention drift impairs the precise visual–textual alignments learned during base training, limiting the model’s ability to focus on relevant regions for accurate object localization. To overcome these limitations, we propose Protoalign, a novel prototype-guided multilayer alignment network that maintains robust cross-modal relationships across multiple network layers while mitigating catastrophic forgetting. Specifically, cross-modal prototype guidance (CPG) enables stable feature learning through class-specific prototype fusion to mitigate feature space distortion. Multimodal feature aggregation (MFA) strengthens feature relationships through channel-level interactions to overcome attention drift. Moreover, we propose an integrator that facilitates consistent information flow between network layers. Protoalign progressively refines multimodal features, enabling effective novel class adaptation while preserving crucial base knowledge. The refined representations are then processed by a detection transformer (DETR) decoder for accurate object detection. Extensive experiments on iSAID, DIOR, FAIR1M-Airplane, and NWPU VHR-10 datasets demonstrate that Protoalign achieves superior performance while significantly reducing catastrophic forgetting compared to existing methods.
- 10 July 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 63 (01962892) , 1-23
Prototype-Guided Multilayer Alignment Network for Few-Shot Object Detection in Remote Sensing
Few-shot object detection (FSOD) addresses the challenge of limited labeled data by enabling detectors to learn from minimal annotations. Recent work on image–text fusion has shown promise in overcoming data scarcity issues. However, these models suffer from catastrophic forgetting due to two distinct challenges during few-shot adaptation. First, feature space distortion disrupts established relationships between image–text modalities and blurs decision boundaries, leading to inaccuracies in object classification. Second, attention drift impairs the precise visual–textual alignments learned during base training, limiting the model’s ability to focus on relevant regions for accurate object localization. To overcome these limitations, we propose Protoalign, a novel prototype-guided multilayer alignment network that maintains robust cross-modal relationships across multiple network layers while mitigating catastrophic forgetting. Specifically, cross-modal prototype guidance (CPG) enables stable feature learning through class-specific prototype fusion to mitigate feature space distortion. Multimodal feature aggregation (MFA) strengthens feature relationships through channel-level interactions to overcome attention drift. Moreover, we propose an integrator that facilitates consistent information flow between network layers. Protoalign progressively refines multimodal features, enabling effective novel class adaptation while preserving crucial base knowledge. The refined representations are then processed by a detection transformer (DETR) decoder for accurate object detection. Extensive experiments on iSAID, DIOR, FAIR1M-Airplane, and NWPU VHR-10 datasets demonstrate that Protoalign achieves superior performance while significantly reducing catastrophic forgetting compared to existing methods.
- 20 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 63 (01962892) , 1-18
Dynamic Adaptive Region Transformer for Tiny-Object Detection in Remote Sensing
Despite DETR-like methods having improved end-to-end detection capabilities, they fundamentally struggle with the uniform processing of entire flattened feature maps, causing queries to attend to irrelevant regions. This results in redundant attention patterns, leading to computational burdens and inefficiencies in the detection of tiny objects in remote sensing imagery across dense and sparse settings. To address these limitations, we present a lightweight Transformer-based encoder-decoder architecture called Dynamic Adaptive Region Transformer (DART). Specifically, Density Adaptive Transformer (DAT) employs an adaptive-region attention (ARA) mechanism that dynamically generates content-aware spatial regions rooted in feature density and semantic relevance. This strategy prioritizes computational resources on semantically rich areas while minimizing focus on irrelevant background regions. Region Aware Decoder (RAD) incorporates a masked region-aware cross-attention (MRA) mechanism, where queries interact exclusively with the adaptive masked regions generated by the encoder, thereby reducing redundant focus on overlapping or irrelevant areas. Meanwhile, a query diversity loss is introduced to penalize overlapping attention patterns among queries, encouraging each query to focus on distinct and complementary regions. By adapting to data density and directing queries to essential areas within the image, DART enhances feature extraction and object localization for various-sized objects in dense and sparse settings. Experimental results demonstrate that DART achieves state-of-the-art performance on AI-TOD, DOTA-v2.0 and LEVIR-Ship benchmarks and exhibits strong generalization capabilities on DIOR, while using only 13 million parameters and a computational cost of 68 GFLOPs.
- 20 June 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 63 (01962892) , 1-18
Dynamic Adaptive Region Transformer for Tiny-Object Detection in Remote Sensing
Despite DETR-like methods having improved end-to-end detection capabilities, they fundamentally struggle with the uniform processing of entire flattened feature maps, causing queries to attend to irrelevant regions. This results in redundant attention patterns, leading to computational burdens and inefficiencies in the detection of tiny objects in remote sensing imagery across dense and sparse settings. To address these limitations, we present a lightweight Transformer-based encoder-decoder architecture called Dynamic Adaptive Region Transformer (DART). Specifically, Density Adaptive Transformer (DAT) employs an adaptive-region attention (ARA) mechanism that dynamically generates content-aware spatial regions rooted in feature density and semantic relevance. This strategy prioritizes computational resources on semantically rich areas while minimizing focus on irrelevant background regions. Region Aware Decoder (RAD) incorporates a masked region-aware cross-attention (MRA) mechanism, where queries interact exclusively with the adaptive masked regions generated by the encoder, thereby reducing redundant focus on overlapping or irrelevant areas. Meanwhile, a query diversity loss is introduced to penalize overlapping attention patterns among queries, encouraging each query to focus on distinct and complementary regions. By adapting to data density and directing queries to essential areas within the image, DART enhances feature extraction and object localization for various-sized objects in dense and sparse settings. Experimental results demonstrate that DART achieves state-of-the-art performance on AI-TOD, DOTA-v2.0 and LEVIR-Ship benchmarks and exhibits strong generalization capabilities on DIOR, while using only 13 million parameters and a computational cost of 68 GFLOPs.
- 21 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 63 (01962892) , 1-21
Memory-Augmented Detection Transformer for Few-Shot Object Detection in Remote Sensing Imagery
The few-shot object detection (FSOD) in remote sensing (RS) faces a fundamental challenge in balancing contextual feature learning with the representation stability. Current approaches either excel at capturing rich contextual relationships through multimodal architectures but struggle with catastrophic forgetting or maintain stable feature representations through incremental learning frameworks at the cost of contextual understanding. This limitation is particularly acute in RS, where complex spatial-contextual dependencies form distinctive co-occurrence patterns important for accurate object detection. We present a memory-augmented detection transformer (MemDeT) that bridges this gap through three key modules: 1) a contextual layerwise fusion (CLF) module that progressively integrates visual and textual information across transformer layers through an adaptive attention mechanism and unified cross-modal fusion, enabling both fine-grained object feature extraction at lower layers and abstract contextual relationship learning at higher layers; 2) a unified episodic memory (UEM) that serves as a dynamic knowledge repository, employing similarity-surprise-based update mechanisms with structured key-value memory organization to strategically retrieve and update relevant past experiences while preserving base knowledge through contextual scoring and dynamic retrieval; and 3) a memory-augmented decoder (MAD) that generates context-aware queries by combining current visual observations with accumulated contextual knowledge through memory-prioritized attention and progressive query refinement. Extensive experiments on RS datasets, NWPU VHR-10, DIOR, and iSAID, demonstrate that MemDeT significantly outperforms state-of-the-art methods in both contextual understanding and knowledge retention. Cross-dataset evaluation further validates MemDeT’s robust generalization capabilities, successfully transferring knowledge from DIOR to iSAID’s urban scenes, particularly in scenarios requiring strong contextual understanding with limited training examples.
- 21 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 63 (01962892) , 1-21
Memory-Augmented Detection Transformer for Few-Shot Object Detection in Remote Sensing Imagery
The few-shot object detection (FSOD) in remote sensing (RS) faces a fundamental challenge in balancing contextual feature learning with the representation stability. Current approaches either excel at capturing rich contextual relationships through multimodal architectures but struggle with catastrophic forgetting or maintain stable feature representations through incremental learning frameworks at the cost of contextual understanding. This limitation is particularly acute in RS, where complex spatial-contextual dependencies form distinctive co-occurrence patterns important for accurate object detection. We present a memory-augmented detection transformer (MemDeT) that bridges this gap through three key modules: 1) a contextual layerwise fusion (CLF) module that progressively integrates visual and textual information across transformer layers through an adaptive attention mechanism and unified cross-modal fusion, enabling both fine-grained object feature extraction at lower layers and abstract contextual relationship learning at higher layers; 2) a unified episodic memory (UEM) that serves as a dynamic knowledge repository, employing similarity-surprise-based update mechanisms with structured key-value memory organization to strategically retrieve and update relevant past experiences while preserving base knowledge through contextual scoring and dynamic retrieval; and 3) a memory-augmented decoder (MAD) that generates context-aware queries by combining current visual observations with accumulated contextual knowledge through memory-prioritized attention and progressive query refinement. Extensive experiments on RS datasets, NWPU VHR-10, DIOR, and iSAID, demonstrate that MemDeT significantly outperforms state-of-the-art methods in both contextual understanding and knowledge retention. Cross-dataset evaluation further validates MemDeT’s robust generalization capabilities, successfully transferring knowledge from DIOR to iSAID’s urban scenes, particularly in scenarios requiring strong contextual understanding with limited training examples.
- 18 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 25 (11) , 19896-19908
Training Sample Selection Method Based on Model-Data-Driven Maritime Radar Target Detection
Multidimensional feature extraction and fusion algorithms are widely utilized to improve the performance of radar maritime target detection. However, these methods often suffer from underfitting and overfitting issues due to the uneven distribution of training samples. To enhance the performance of maritime target detection while mitigating the issues of underfitting and overfitting, an effective approach is to improve the comprehensiveness of training samples. This study proposes a novel training sample selection method based on model-data-driven maritime radar target detection. A surrogate equation is constructed in this study using existing empirical models and measured data. Subsequently, it dynamically selects training samples exhibiting strong coverage and compactness, guided by the global and local variation characteristics of this surrogate equation. Experimental results show that, when using the same feature and classifier conditions, the feature classifiers formed by the proposed method outperform those built using traditional sample selection approaches (such as random assignment, expert knowledge, and statistical methods). This superiority is evident across various target types, including daily vessels, small fishing boats, and marine buoys, as well as under diverse sea conditions and observation parameters like incidence angles. This approach not only enhances the performance of maritime target detection, but also ensures the generalization ability of the method.
- 18 April 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Sensors Journal
- Vol. 25 (11) , 19896-19908
Training Sample Selection Method Based on Model-Data-Driven Maritime Radar Target Detection
Multidimensional feature extraction and fusion algorithms are widely utilized to improve the performance of radar maritime target detection. However, these methods often suffer from underfitting and overfitting issues due to the uneven distribution of training samples. To enhance the performance of maritime target detection while mitigating the issues of underfitting and overfitting, an effective approach is to improve the comprehensiveness of training samples. This study proposes a novel training sample selection method based on model-data-driven maritime radar target detection. A surrogate equation is constructed in this study using existing empirical models and measured data. Subsequently, it dynamically selects training samples exhibiting strong coverage and compactness, guided by the global and local variation characteristics of this surrogate equation. Experimental results show that, when using the same feature and classifier conditions, the feature classifiers formed by the proposed method outperform those built using traditional sample selection approaches (such as random assignment, expert knowledge, and statistical methods). This superiority is evident across various target types, including daily vessels, small fishing boats, and marine buoys, as well as under diverse sea conditions and observation parameters like incidence angles. This approach not only enhances the performance of maritime target detection, but also ensures the generalization ability of the method.
- 17 April 2025
- Published by Elsevier in Expert Systems with Applications
- Vol. 288
Dual-Scale Former with Fourier Feature Pyramid Network for Tiny Object Detection in Remote Sensing
- 17 April 2025
- Published by Elsevier in Expert Systems with Applications
- Vol. 288
Dual-Scale Former with Fourier Feature Pyramid Network for Tiny Object Detection in Remote Sensing
- 20 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 63 (01962892) , 1-15
Infrared Small Target Detection Based on Interpretation Weighted Sparse Method
Accurately and robustly detecting infrared (IR) small targets plays an essential role in infrared search and tracking (IRST) systems against complex backgrounds. However, the low signal-to-clutter ratio (SCR) feature target, also known as the dim target, is more similar to the local background region, making it prone to being misdetected as background or buried in clutter interferences, ultimately leading to detection failure. This has become a common and unavoidable challenge in the field. To alleviate this issue, an interpretation weighted sparse (IWS) method is proposed to detect IR small and dim targets while maintaining clutter suppression capability. The interpretability operation of the proposed method stems from an analysis of IR target characteristics, including dim target feature extraction and clutter suppression. Based on the multi-subspaces model, the relevant feature of the dim target is mined by a designed structure tensor, improving the target’s sparsity. Then, the contrast characteristic is introduced to eliminate the cost of dim target mine operation, namely residual clutter interference. Finally, the IWS map is obtained by fusing the target detection module and clutter suppression module, after which an adaptive threshold segmentation operation is adopted to extract the small targets. A series of experimental results and application discussion cases demonstrate that IWS outperforms other existing methods regarding detection performance while also exhibiting background suppression ability.
- 20 January 2025
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Transactions on Geoscience and Remote Sensing
- Vol. 63 (01962892) , 1-15
Infrared Small Target Detection Based on Interpretation Weighted Sparse Method
Accurately and robustly detecting infrared (IR) small targets plays an essential role in infrared search and tracking (IRST) systems against complex backgrounds. However, the low signal-to-clutter ratio (SCR) feature target, also known as the dim target, is more similar to the local background region, making it prone to being misdetected as background or buried in clutter interferences, ultimately leading to detection failure. This has become a common and unavoidable challenge in the field. To alleviate this issue, an interpretation weighted sparse (IWS) method is proposed to detect IR small and dim targets while maintaining clutter suppression capability. The interpretability operation of the proposed method stems from an analysis of IR target characteristics, including dim target feature extraction and clutter suppression. Based on the multi-subspaces model, the relevant feature of the dim target is mined by a designed structure tensor, improving the target’s sparsity. Then, the contrast characteristic is introduced to eliminate the cost of dim target mine operation, namely residual clutter interference. Finally, the IWS map is obtained by fusing the target detection module and clutter suppression module, after which an adaptive threshold segmentation operation is adopted to extract the small targets. A series of experimental results and application discussion cases demonstrate that IWS outperforms other existing methods regarding detection performance while also exhibiting background suppression ability.
- Open Access
- 2 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Vol. 17 (19391404) , 16005-16019
A Real-Time Infrared Small Target Detection Based on Double Dilate Contrast Measure
The accurate and robust detection of small infrared (IR) targets in complex backgrounds is crucial for the effective operation of IR search and track (IRST) systems. However, the high-intensity components in background regions could easily be confused with the real targets among local contrast features, which makes it an inevitable and challenging problem. To efficiently alleviate this issue, a real-time infrared small target detection method is proposed by using the double dilate contrast measure (DDCM). The proposed method is a robust and efficient detector that takes into account various aspects, including target property, background information, and their interrelations. First, an enhancement layer is considered to enlarge the difference between the target and the background with dilation operation as the core component. Second, based on considering the relationship between the local maximum and regional mean, a suppression layer is introduced to further eliminate the clutter and noise from the high-intensity regions, and the dilate operation is also used as the core component. Finally, the DDCM map is obtained by fusing the enhancement and suppression layers, after which an adaptive segmentation operation is adopted to extract the small targets. Extensive experimental results on nine sequences demonstrate that DDCM outperforms other existing methods in terms of detection rate and false alarm rate while also exhibiting high efficiency.
- 2 July 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Vol. 17 (19391404) , 16005-16019
Open AccessA Real-Time Infrared Small Target Detection Based on Double Dilate Contrast Measure
The accurate and robust detection of small infrared (IR) targets in complex backgrounds is crucial for the effective operation of IR search and track (IRST) systems. However, the high-intensity components in background regions could easily be confused with the real targets among local contrast features, which makes it an inevitable and challenging problem. To efficiently alleviate this issue, a real-time infrared small target detection method is proposed by using the double dilate contrast measure (DDCM). The proposed method is a robust and efficient detector that takes into account various aspects, including target property, background information, and their interrelations. First, an enhancement layer is considered to enlarge the difference between the target and the background with dilation operation as the core component. Second, based on considering the relationship between the local maximum and regional mean, a suppression layer is introduced to further eliminate the clutter and noise from the high-intensity regions, and the dilate operation is also used as the core component. Finally, the DDCM map is obtained by fusing the enhancement and suppression layers, after which an adaptive segmentation operation is adopted to extract the small targets. Extensive experimental results on nine sequences demonstrate that DDCM outperforms other existing methods in terms of detection rate and false alarm rate while also exhibiting high efficiency.
- 24 June 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Letters
- Vol. 21 (1545598X) , 1-5
Multidirectional Diffusion Difference Measure Approach for IR Small Target Detection
The accurate detection of small infrared (IR) targets in complex backgrounds is crucial for the effective operation of IR search and track (IRST) systems. In this letter, we propose a novel multi-directional diffusion difference measure (MDDDM) approach for accurate and low-complexity IR small target detection. MDDDM consists of three stages. Firstly, a gray intensity measure (GIM) is generated to highlight the important features and suppress high-intensity background regions by using a computationally efficient difference of Gaussian (DoG) filter. Secondly, the diffusion difference measure (DDM) kernel is computed by fusing an improved Gaussian kernel (IGK) and a directional filter kernel (DFK), utilizing feature minimization instead of omnidirectional gradient judgment to suppress background edge clutters. Lastly, threshold segmentation operation is used for precise localization and extraction of the target from the background. Experimental results demonstrate that MDDDM outperforms other state-of-the-art algorithms in terms of detection rate and false alarm rate while maintaining the fastest running speed.
- 24 June 2024
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Geoscience and Remote Sensing Letters
- Vol. 21 (1545598X) , 1-5
Multidirectional Diffusion Difference Measure Approach for IR Small Target Detection
The accurate detection of small infrared (IR) targets in complex backgrounds is crucial for the effective operation of IR search and track (IRST) systems. In this letter, we propose a novel multi-directional diffusion difference measure (MDDDM) approach for accurate and low-complexity IR small target detection. MDDDM consists of three stages. Firstly, a gray intensity measure (GIM) is generated to highlight the important features and suppress high-intensity background regions by using a computationally efficient difference of Gaussian (DoG) filter. Secondly, the diffusion difference measure (DDM) kernel is computed by fusing an improved Gaussian kernel (IGK) and a directional filter kernel (DFK), utilizing feature minimization instead of omnidirectional gradient judgment to suppress background edge clutters. Lastly, threshold segmentation operation is used for precise localization and extraction of the target from the background. Experimental results demonstrate that MDDDM outperforms other state-of-the-art algorithms in terms of detection rate and false alarm rate while maintaining the fastest running speed.
- Open Access
- 30 August 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Vol. 16 (19391404) , 8475-8491
Multiscale Context-Aware Feature Fusion Network for Land-Cover Classification of Urban Scene Imagery
Recently, several land cover classification models have achieved great success in terms of both accuracy and computational performance. However, it remains challenging due to inter-class similarities, intra-class variations, scale-related inaccuracies, and high computational complexity. First, these methods fail to establish a correlation among different feature maps during multiscale feature extraction, leading to inter-class similarities and intra-class variations. Second, they under-utilize feature interdependencies of the context contained in each layer of the encoder-decoder architecture, causing scale-related inaccuracies. Third, they cause checkerboard artifacts and blurry edges, which can negatively impact the accuracy and generated segmentation map at increased computational cost. To address these problems, this article proposes a novel multiscale context-aware feature fusion network (MCN) for high-resolution urban scene images. MCN mainly consists of three modules: 1) multiscale feature enhancement module (MFE) for backbone network to extract rich spatial information dynamically by incorporating dense correlation among feature maps with different receptive fields, 2) multilayer feature fusion module (MLF) as skip connections to produce a single high-level representation of the local-global context by capturing low-, mid-, high- level interdependencies at different encoder-decoder stages, and 3) pixel shuffle decoder (PSD) to reduce the blurry edges and checkerboard artifacts while upsampling with reduced number of parameters. Experiments on three high-resolution aerial and satellite urban scene datasets show that MCN consistently outperforms the mainstream land cover classification models. Specifically, MCN achieves an OA of 93.51 on Potsdam, 90.18 on Vaihingen, and a mIoU of 73.73 on DeepGlobe.
- 30 August 2023
- Published by Institute of Electrical and Electronics Engineers (IEEE) in IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing
- Vol. 16 (19391404) , 8475-8491
Open AccessMultiscale Context-Aware Feature Fusion Network for Land-Cover Classification of Urban Scene Imagery
Recently, several land cover classification models have achieved great success in terms of both accuracy and computational performance. However, it remains challenging due to inter-class similarities, intra-class variations, scale-related inaccuracies, and high computational complexity. First, these methods fail to establish a correlation among different feature maps during multiscale feature extraction, leading to inter-class similarities and intra-class variations. Second, they under-utilize feature interdependencies of the context contained in each layer of the encoder-decoder architecture, causing scale-related inaccuracies. Third, they cause checkerboard artifacts and blurry edges, which can negatively impact the accuracy and generated segmentation map at increased computational cost. To address these problems, this article proposes a novel multiscale context-aware feature fusion network (MCN) for high-resolution urban scene images. MCN mainly consists of three modules: 1) multiscale feature enhancement module (MFE) for backbone network to extract rich spatial information dynamically by incorporating dense correlation among feature maps with different receptive fields, 2) multilayer feature fusion module (MLF) as skip connections to produce a single high-level representation of the local-global context by capturing low-, mid-, high- level interdependencies at different encoder-decoder stages, and 3) pixel shuffle decoder (PSD) to reduce the blurry edges and checkerboard artifacts while upsampling with reduced number of parameters. Experiments on three high-resolution aerial and satellite urban scene datasets show that MCN consistently outperforms the mainstream land cover classification models. Specifically, MCN achieves an OA of 93.51 on Potsdam, 90.18 on Vaihingen, and a mIoU of 73.73 on DeepGlobe.