Model design and training
Model design must balance predictive accuracy against inference throughput at continental or global scale, and accommodate the heterogeneity of EO data. Effective model training must examine the specificities of large EO datasets to guide hyperparameter tuning and data augmentation.
Trade-off
Developing ML models for large-scale EO applications requires balancing model complexity, predictive performance, and inference throughput. This trade-off becomes critical in operational settings, where maps must be produced on fixed schedules (quarterly to yearly monitoring) or on demand (real-time alerts, post-disaster response) over wide areas. It is largely shaped by two choices: the model family, and the use of spatial context. Model families range widely in cost. At the simple end, tree ensembles such as gradient boosting and random forests remain common in operational products [1], [2], [3], being cheap to train and deploy without GPUs. At the complex end, Vision Transformer architectures have attracted growing interest, including for large-area mapping [4], [5], [6], [7]. Simpler convolutional encoder-decoder architectures remain widely adopted in operational pipelines, where the latency and memory constraints often favor their lower computational footprint [8], [9], [10]. In particular, UNet architectures [11] with EfficientNet backbones [12] offer a well-documented trade-off between accuracy and throughput for such settings. Most of these architectures operate on image patches, exploiting the spatial context around each pixel. Many operational products instead predict each pixel independently, from its own time series [1], [4], [13], giving up spatial information in exchange for smaller inputs and simpler inference. Knowledge distillation [14] offers a further avenue to improve inference efficiency without notable performance degradation [15].
A recent benchmark of inference throughput across 45 vision backbones and geospatial foundation-model encoders illustrates the stakes: on a single H100 GPU, forward-pass throughput varies by roughly 205× between the fastest and slowest models, whereas moving the same model from a V100 to an H100 yields only a 3–5× speedup. For planetary-scale inference, model choice therefore dominates hardware choice, and combining half precision with torch.compile gives the largest gains [16].
Handling input heterogeneity
EO applications frequently combine inputs that differ in spatial resolution, number of channels, or temporal cadence. Rather than resampling all modalities to a common grid before ingestion (as described in From grid to patches), an alternative approach consists in delegating this alignment to the model itself. Typical designs employ modality-specific encoder branches that operate at each source’s native resolution and fuse representations at a shared spatial scale, either through learned upsampling, cross-attention, or feature concatenation after spatial alignment layers [17]. This strategy avoids the interpolation artifacts and information loss inherent in a priori resampling. Recent geospatial foundation models have pushed this flexibility further, using dynamic patch embeddings and scale-adaptive encoders to train a single model across sensors with arbitrary channel counts, resolutions, and tile extents, while also gracefully handling missing or variable modalities at inference time [18], [19], [20]. The added architectural complexity must, however, be weighed against the throughput and simplicity advantages of a single pre-aligned input tensor, particularly in operational pipelines where inference latency is a binding constraint.
Hyperparameter tuning
Hyperparameter optimization (HPO) is an important step in the ML pipeline, and we refer the reader to [21] for a general review of HPO good practices. We focus on two key aspects that are relevant for large EO datasets: which parameters to optimize, and how much data to use for it. Which parameters matter depends on the model family. Hyperparameter tuning for deep learning models is inherently challenging: the number of configurable hyperparameters is large, and their relative influence on model performance varies across tasks, architectures, and datasets, making it difficult to establish universally valid tuning priorities. Nevertheless, a broad consensus has emerged in the literature that the learning rate and its scheduler, batch size, and regularization mechanisms such as dropout and weight decay tend to be the most influential hyperparameters for deep neural networks [22]. Classical ML models have fewer hyperparameters: for gradient boosting trees, for example, the learning rate is the main one to tune [23]. Regarding how much data should be used for HPO, the common practice is to use the entire validation set, as in [4], [24]. However, given the computational cost of training on large datasets, a strategy encountered in the literature (e.g., [25]) is to perform hyperparameter optimization on a subset of the data, and subsequently train the final model on the full dataset using the selected configuration. This approach is a special case of multi-fidelity hyperparameter optimization, and many methods have been developed for it [26], [27], [28], ranging from simple random subsampling to gradient-based subset selection. When complex subset selection methods are not warranted, stratified random sampling should be preferred over naive uniform sampling, particularly in the presence of class imbalance or long-tailed distributions, to ensure that the subset adequately reflects the class distribution of the original dataset. When using the full dataset is prohibitively expensive, we suggest adopting a two-stage strategy: first using the subset for efficient hyperparameter search to narrow the configuration space, then validation of the top-performing configurations on the full dataset before final model training. We encourage researchers to document their hyperparameter optimization or to report when none was performed, as in [10].
Data augmentation
In the SatML field, less attention is brought to data augmentation than in the general ML field. This is partly due to the fact that “augmentation techniques designed for natural images should not be applied to satellite imagery without careful consideration” [29]. Their findings suggest that while standard natural image techniques (particularly color-based augmentations) do not translate effectively to the satellite domain, geometric operations remain a beneficial approach. Furthermore, Muhawenayo et al. [30] have identified that without specific data augmentation techniques to make them resilient, ML models often lack the necessary robustness to handle domain shifts (e.g., in terms of brightness) which can lead to decreased performance and visual artifacts. Various data augmentation techniques have been established in the literature to improve model generalization. Geometric transformations (such as rotations, flips, and random cropping) are being applied to EO data [8], [9], [31], [32], although particular care must be taken for SAR, since the look direction fixes which side of a slope shows layover or shadow. Recent research has introduced domain-specific augmentations tailored to EO data. To enhance spectral robustness, noise is frequently added to Sentinel-2 surface reflectance values [4], [31], while Muhawenayo et al. [30] advocate for varying Sentinel-2 processing strategies to mitigate sensitivity to shifting processing conventions. For models incorporating spatial metadata, Van Tricht et al. [1] apply random perturbations to latitude and longitude coordinates. Furthermore, temporal augmentations, such as dropping timesteps, are utilized to simulate cloud-induced data gaps [4] or missing modalities. Finally, several approaches specifically target sensor-induced artifacts to ensure resilience during inference; these include simulating narrow no-data seams in Sentinel-1 GRD mosaics [33], as well as modeling swath boundaries, misregistration, and detector failures [34].
A model trained with the practices described above will produce a prediction for every pixel on Earth. However, a prediction should come with a measure of its confidence. The model cannot perform equally well everywhere: it will be more reliable in regions well-represented in the training data than in under-sampled biomes, more confident under clear-sky conditions than under persistent cloud cover, and more accurate for common classes than for rare ones. Characterizing where and how much to trust the predictions is the subject of the next section, which surveys ML uncertainty quantification methods and argues that uncertainty is not a supplementary layer to be added as an afterthought, but a necessary component of any credible map product.