ResNet

ResNet, or Residual Network, is a deep learning architecture introduced by Microsoft Research that solved the degradation problem in very deep neural networks through residual learning. Instead of learning a direct mapping from input to output, each residual block learns the difference (residual) between the desired output and the input, which is then added back via a skip connection. This simple but powerful innovation enables training of networks with hundreds or thousands of layers, far deeper than was previously possible, by ensuring that gradients flow freely through the network during backpropagationBackpropagationBackpropagation is the fundamental algorithm for computing gradients in neural network training, propagating error si.... ResNet as a Backbone for Geospatial ModelsResNet architectures, particularly ResNet-50 and ResNet-101, serve as the backbone feature extractors in many geospatial AI systems. Object detectionObject DetectionObject Detection is a computer vision technique that identifies and localizes specific objects within images or video... frameworks like Faster R-CNN and RetinaNet use ResNet to extract multi-scale features from satellite imagerySatellite ImagerySatellite imagery consists of photographs and data captured by Earth observation satellites orbiting the planet. Thes... for building and vehicle detection. Semantic segmentationSemantic SegmentationSemantic Segmentation is a computer vision technique that assigns a class label to every pixel in an image, enabling ... models like DeepLab and Feature Pyramid Networks build upon ResNet backbones to produce pixel-level land cover maps. Change detectionChange DetectionChange detection uses geospatial data and imagery to track and analyze alterations in landscapes, infrastructure, or ... systems use dual ResNet encoders to extract comparable features from multi-temporal image pairs. Pretrained ResNet models, initially trained on ImageNet, transfer effectively to remote sensingRemote SensingRemote sensing is the science of collecting data about Earth's surface without direct physical contact, primarily usi... tasks through fine-tuning, reducing the need for large labeled geospatial datasets. Variants and Continued RelevanceResNet variants include ResNeXt with grouped convolutions for improved efficiency, Wide ResNets with wider layers, and SE-ResNet with channel attention. While Vision Transformers have surpassed ResNet on some benchmarks, ResNet remains widely used due to its efficiency, well-understood behavior, and extensive ecosystem of pretrained weights. ResNet's architecture directly inspired the skip connections in U-NetU-NetU-Net is an encoder-decoder neural network architecture with skip connections designed for precise image segmentation... and other segmentation networks that are central to geospatial image analysis.