Dropout

Dropout is a regularizationRegularizationRegularization encompasses techniques that prevent machine learning models from overfitting to training data, ensurin... method for neural networks where randomly selected neurons are temporarily removed, along with all their connections, during each training step. Each neuron is dropped with a specified probability, typically 20-50%, meaning different subnetworks are trained on different mini-batches. This forces the network to develop redundant representations and prevents co-adaptation, where neurons become overly dependent on specific other neurons. At inference time, all neurons are active but their outputs are scaled by the dropout probability to maintain consistent expected values. Preventing Overfitting in Geospatial ModelsOverfitting is a persistent challenge in geospatial AI because training data often covers limited geographic regions, time periods, or sensor conditions, while models must generalize to diverse real-world scenarios. A model trained on satellite imagerySatellite ImagerySatellite imagery consists of photographs and data captured by Earth observation satellites orbiting the planet. Thes... from one region may memorize region-specific patterns rather than learning generalizable features. Dropout mitigates this by ensuring that the model cannot rely on any single feature or neuron, building more robust representations that transfer better across geographic contexts. It is commonly applied in the fully connected layers of classification networks and in the decoder paths of segmentation architectures like U-NetU-NetU-Net is an encoder-decoder neural network architecture with skip connections designed for precise image segmentation.... Practical Usage and AlternativesDropout rates are a tunable hyperparameter, with higher rates providing stronger regularization but potentially limiting model capacity. Spatial Dropout, which drops entire feature map channels rather than individual neurons, is often more effective for convolutional architectures processing satellite imagery because adjacent pixels are highly correlated. DropBlock extends this concept by dropping contiguous regions of feature maps. Modern architectures sometimes replace dropout with other regularization techniques like Batch NormalizationBatch NormalizationBatch Normalization is a technique that normalizes the inputs to each layer during neural network training, stabilizi..., weight decay, or data augmentationData AugmentationData Augmentation expands training datasets through transformations like rotation, flipping, color shifting, and crop..., though combining multiple regularization strategies often produces the best results for geospatial tasks.