Land use land cover (LULC) mapping using deep learning greatly enhances our understanding of geographic, socioeconomic, and urban development patterns. However, annotated satellite data are scarce in developing countries in South and East Asia due to funding and urban heterogeneity. We present the Bangladesh Open LULC Map (BOLM), a high-resolution dataset with pixel-level annotations across eleven classes for the Dhaka metropolitan area and surrounding regions (4,392 \(\hbox {km}^2\) , 891 million pixels). Annotations were generated using high-resolution Bing imagery following a rigorous, multi-stage validation process supported by GIS experts. Although Bing data provides fine spatial detail for visual interpretation, multispectral data remain essential for spectral analysis and temporal monitoring. We benchmarked DeepLabV3+, HRNetv2, U-net, UnimatchV2 and Segmenter ViT-16 on five merged LULC classes, achieving an overall IoU of 0.50 and F1 score of 0.66 with the UnimatchV2. Further experiments using Sentinel-2A data highlight the trade-off between spatial and spectral resolution. Our results show that while DeepLabV3+ and HRNet are effective for LULC mapping, other models such as U-Net, Segmenter, and UniMatchV2 perform even better, highlighting the model-agnostic nature of our study. The BOLM dataset can support future deep learning models and domain adaptation for LULC tasks in data-scarce regions, addressing the critical gap of high-quality LULC data in South and East Asia.
