When the Rules Became Weights

THE OLD ARRANGEMENT

For a long time, automated Earth analysis meant turning an expert's judgment into instructions. Vegetation could be separated from bare ground with a spectral ratio. Water could be isolated with thresholds. Texture could help distinguish a city from a bright desert. The computer performed the calculation quickly, but the idea of what mattered still came from a person.

That arrangement was more powerful than it now sounds. An index or decision rule is inspectable, portable and often physically interpretable. If the output looks wrong, an analyst can ask which threshold fired and why. But the method becomes awkward when the useful signal is spread across dozens of bands, neighbouring pixels, seasons and contexts that do not reduce cleanly to one rule.

Machine learning changed the division of labour. Instead of specifying the pattern, people supplied examples and an objective. The model adjusted internal parameters until the examples separated. The rules did not disappear. They became millions, and then billions, of weights that no committee was expected to review individually. Civilization accepted this surprisingly quickly.

THE TECHNICAL SHIFT

Feature engineering became feature learning

The intermediate era is important because machine learning did not begin with deep neural networks. Remote-sensing teams used logistic regression, support-vector machines, random forests and other statistical classifiers for years. These systems could learn a decision boundary from labelled examples, but the inputs were still usually designed by people: vegetation indices, band ratios, texture measures, elevation, seasonal statistics and neighbourhood features.

Deep learning moved the learning process one level earlier. A convolutional network can learn filters that respond to edges, shapes, textures and larger spatial arrangements directly from imagery. Later layers combine those responses into structures useful for the task. A road is not only a particular reflectance value. It is narrow, connected and persistent across space. A settlement is not one bright pixel. It is a pattern of roofs, streets, shadows and geometry.

That matters because Earth categories are rarely defined by one measurement. Bright rock can resemble concrete. A dry riverbed can resemble a road. A recently harvested field can resemble bare ground. Spatial context lets a model use relationships that would be tedious to encode as rules and brittle to maintain across continents.

The price is interpretability. In a hand-designed system, the feature has a name before the model uses it. In a learned system, a useful internal feature may have no human label at all. It exists because changing that pattern helped reduce the training error. The system gains expressive power and loses the comforting fiction that every useful distinction has already been named.

DYNAMIC WORLD

A global classifier that keeps updating

Dynamic World shows what this shift looks like when it becomes infrastructure. Developed by Google and the World Resources Institute, it applies a fully convolutional neural network to Sentinel-2 imagery and produces near-real-time land-cover predictions at 10-meter resolution. The taxonomy has nine classes, including water, trees, grass, flooded vegetation, crops, shrub and scrub, built area, bare ground, and snow or ice.

The training process still depended heavily on people. The project assembled roughly 20,000 hand-labelled image tiles distributed across global biomes and regions. Annotators used higher-resolution imagery and other reference information to decide what the pixels represented. Learning reduced the need to write the classification rules, not the need to define the categories or provide examples of them.

Once trained, the model could be applied continuously as new Sentinel-2 scenes arrived. That changed the product from a map published occasionally into a stream of probabilities tied to individual acquisitions. A place could be classified again whenever a usable scene appeared, which is a different operational object from an annual land-cover layer even when the colour palette looks familiar.

Dynamic World also exposes a probability for every class rather than only the winning label. A location may look mostly like crops but retain substantial probability for grass or bare ground. That ambiguity is not an inconvenience to be hidden. It is information about what the model actually saw, and later chapters will need it when several systems begin passing beliefs to one another.

DISTRIBUTION SHIFT

The model also learns where its training data came from

A learned feature is useful only to the extent that the training examples taught the right regularity. This becomes a geographic problem very quickly. Roof materials, field sizes, irrigation patterns, seasonal calendars, cloud regimes and sensor conditions differ by region. A model can learn a feature that is robust across its validation set and discover a new planet when deployed two countries away.

The Dynamic World paper reports exactly this kind of variation. Stable classes such as water and trees tend to be easier, while mixed or transient classes such as shrub and scrub, bare ground, crops, grass and flooded vegetation are harder. Missed cloud and shadow can also become confident land-cover errors. The model is global. Its uncertainty is not evenly distributed across the globe.

The WILDS benchmark made the broader problem explicit by testing models under real distribution shifts, including satellite imagery separated across geography and time. Standard models perform substantially worse out of distribution than on data resembling their training set. Random train-test splits can therefore produce a reassuring score for a deployment condition nobody will actually encounter.

This is one reason planetary AI cannot be judged only by average accuracy. Geography is not noise around the dataset. Geography is part of the dataset-generating process. A model that works in aggregate can still be systematically weak in the exact landscapes where observations, labels and institutions are already sparse.

WHAT CHANGED, AND WHAT DID NOT

Learning a pattern is still not learning a mechanism

The move from rules to weights solved a specific problem. It made it possible to recognize patterns too complicated to specify one feature at a time and to apply that recognition at scales where hand-built rule sets became unmanageable. It did not give the model a physical account of why the pattern exists.

A classifier can learn that a combination of texture, geometry and reflectance usually means cropland without representing planting decisions, irrigation, soil or markets. That is not a defect when the task is land cover. It becomes a defect only when somebody asks the output to carry an explanation it was never trained to contain.

The larger step comes next. A static classifier treats each observation as a fresh problem. Earth is rarely so polite. Fields have seasons, reservoirs fill and empty, construction sites evolve, coastlines move and forests recover. If intelligence is going to mean more than recognition, the model has to learn that the order of observations matters.

The first machine-learned Earth was a collection of patterns. The next one had to remember which pattern came before which.