Abstract Accurate classification of urban land cover is essential for sustainable development and environmental planning, particularly in rapidly urbanizing regions like Delhi. This study proposes UrbanCA-Net, a lightweight 1D Convolutional Neural Network enhanced with Squeeze-and-Excitation channel attention and trained using Focal Loss function to address class imbalance. The model utilized total 19 spectral features (bands and derived indices) from Sentinel-2 Level-2A satellite imagery, for pixel-wise classification into four land cover types: agriculture, water bodies, built-up, and vegetation. UrbanCA-Net achieves a maximum accuracy of 97.98%, with an average macro F1-score of 94.10% across districts, outperforming other baseline CNN models. This study conducted district-wise land cover change analysis across all 11 districts of Delhi for the years 2019 and 2024. Over the study years, a land cover change analysis has shown exceptional increase in built-up and vegetation, while remarkable decrease in agricultural land for South West and North West Delhi. Water bodies remain stable throughout Delhi with notable decrease in North West Delhi. Overall results confirm significant land use transitions throughout Delhi, notably rise in vegetation is a good ecological indicator, the simultaneous decline in agriculture land, and rise in built-up land cover can be harmful for Delhi in future and require sustainable planning. This work can assist urban planners, policy makers, and climate change authorities to mitigate urbanization effects on the environmental factors including elevated temperature, pollution level, and pressure on natural resources. Research highlights <UnorderedList Mark="Bullet"> <ItemContent> UrbanCA-Net: a lightweight 1D-CNN integrated with SE attention for Sentinel-2 LULC mapping. </ItemContent> <ItemContent> Focal Loss mitigates class imbalance, improving macro-F1 performance across districts. </ItemContent> <ItemContent> Delhi-wide LULC (2019–2024): built-up ↑, vegetation ↑, agriculture ↓ — reflecting rapid urban growth. </ItemContent> <ItemContent> Achieves 94.20 % mean accuracy (maximum 98 %), ensuring robust generalization across 11 districts. </ItemContent> <ItemContent> Ablation study confirms that SE + Focal Loss combination outperforms baseline CNN variants. </ItemContent> </UnorderedList>

UrbanCA-Net: a channel attention CNN for urban land cover classification in Delhi using sentinel-2 imagery

期刊论文
发表日期：

2026年4月21日

查看全文

Himanshu Srivastava,
Akansha Singh,
Anuj Kumar Bharti

摘要

Abstract

Accurate classification of urban land cover is essential for sustainable development and environmental planning, particularly in rapidly urbanizing regions like Delhi. This study proposes UrbanCA-Net, a lightweight 1D Convolutional Neural Network enhanced with Squeeze-and-Excitation channel attention and trained using Focal Loss function to address class imbalance. The model utilized total 19 spectral features (bands and derived indices) from Sentinel-2 Level-2A satellite imagery, for pixel-wise classification into four land cover types: agriculture, water bodies, built-up, and vegetation. UrbanCA-Net achieves a maximum accuracy of 97.98%, with an average macro F1-score of 94.10% across districts, outperforming other baseline CNN models. This study conducted district-wise land cover change analysis across all 11 districts of Delhi for the years 2019 and 2024. Over the study years, a land cover change analysis has shown exceptional increase in built-up and vegetation, while remarkable decrease in agricultural land for South West and North West Delhi. Water bodies remain stable throughout Delhi with notable decrease in North West Delhi. Overall results confirm significant land use transitions throughout Delhi, notably rise in vegetation is a good ecological indicator, the simultaneous decline in agriculture land, and rise in built-up land cover can be harmful for Delhi in future and require sustainable planning. This work can assist urban planners, policy makers, and climate change authorities to mitigate urbanization effects on the environmental factors including elevated temperature, pollution level, and pressure on natural resources.

Research highlights

UrbanCA-Net: a lightweight 1D-CNN integrated with SE attention for Sentinel-2 LULC mapping.

Focal Loss mitigates class imbalance, improving macro-F1 performance across districts.

Delhi-wide LULC (2019–2024): built-up ↑, vegetation ↑, agriculture ↓ — reflecting rapid urban growth.

Achieves 94.20 % mean accuracy (maximum 98 %), ensuring robust generalization across 11 districts.

Ablation study confirms that SE + Focal Loss combination outperforms baseline CNN variants.