Wildfires are increasingly recognized as major environmental hazards that endanger ecosystems, human health, and economic stability. In practice, various monitoring efforts, such as hotspot detection, burned-area indices, and field inspections still depend on reliable spatial information to quantify fire impacts over large regions. Accurate segmentation of burned areas from satellite imagery plays a vital role in supporting post-disaster response and sustainable land management. This paper proposes a novel framework for burned area segmentation using multispectral imagery sampled by Random Subspace Band Selection (RSBS) with a U-Net architecture with an ensemble encoder. The RSBS module generates multiple 3-band subsets from Landsat-8 data, incorporating spectrally informative bands such as Near-Infrared (NIR) or Shortwave Infrared (SWIR). These subsets are used to train U-Net models with an ensemble of encoder backbones ResNet34, ResNet50, DenseNet121, MobileNetV2 and InceptionV4 for spectral and architectural diversity. The final predictions are aggregated using majority voting to enhance robustness and generalization. The framework is evaluated on a publicly available Indonesian burned area dataset encompassing diverse land cover types. Experimental results demonstrate that the proposed ensemble model achieves up to 0.7673 IoU, 0.8734 mIoU, and 0.8683 F1 Score, outperforming the best single-backbone model. These findings confirm that the proposed framework offers a scalable, accurate, and adaptable solution for wildfire damage assessment using satellite data.