Machine Learning Map of Brazil’s Caatinga Reaches New Accuracy by Reading Terrain and Climate

Brazil’s Caatinga, the country’s only entirely endemic biome, has just been mapped with an unusual degree of precision, and the secret was not better satellite imagery alone. A research team led by scientists at the Federal University of Rio Grande do Norte and the Federal Rural University of Pernambuco built a land use and land cover classification model for the Caatinga of Rio Grande do Norte that reaches an average overall accuracy of 0.83 and a kappa index of 0.80, figures that outperform the roughly 79.4 percent accuracy reported for the biome in the latest collection of Brazil’s flagship MapBiomas program. Their approach, published in Environmental Monitoring and Assessment, feeds a random forest classifier not only with spectral data but also with rainfall, surface temperature, altitude, and geographic position, letting the algorithm see the landscape the way an ecologist would.

The study area covers approximately 50,746 square kilometers of Rio Grande do Norte, a state where 155 of 167 municipalities fall within the semi-arid domain. Rainfall there is low, irregular, and highly variable from year to year, driven mainly by the seasonal migration of the Intertropical Convergence Zone, while the humid eastern coast receives moisture from Atlantic easterly disturbances and trade winds. This climatic gradient, superimposed on a dramatic geological contrast between ancient crystalline basement and younger sedimentary basins, produces a mosaic of hyperxerophilous and hypoxerophilous Caatinga, Atlantic Forest transitions, mangroves, dunes, and salt flats that has long frustrated automated mapping systems.

To capture that complexity, the researchers segmented the state into six analytical subunits by intersecting hydrographic basin subdivisions with geomorphological domains. The Apodi-Mossoró and Piranhas-Açu basins were each split into crystalline and sedimentary portions, and a cluster of smaller eastern basins was likewise divided, yielding AM Crystalline, AM Sedimentary, PA Crystalline, PA Sedimentary, A3 Crystalline, and A3 Sedimentary units. The logic is physical: sedimentary terrains carry deeper, flatter, more fertile soils suited to mechanized and irrigated agriculture, while crystalline landscapes are shallow, stony, and reddened by iron oxides, favoring open Caatinga and extensive livestock. Splitting the territory along these lines reduced spectral confusion between land use classes and bare soil, a persistent problem when a single classifier must handle an entire heterogeneous biome.

The classification engine itself was the random forest algorithm, an ensemble of decision trees trained by bootstrap aggregation, run on the Google Earth Engine cloud platform. Landsat-8 OLI imagery for 2023 was converted to surface reflectance and filtered for clouds and shadows, then enriched with an extraordinary stack of 153 predictor bands: six spectral bands, dozens of vegetation and water indices, tasseled cap brightness, greenness, and wetness transformations, fraction images from a linear spectral mixture model, land surface temperature derived from the thermal band, and estimated precipitation. The rainfall layer was itself a machine learning product, built from 198 rain gauges and modeled against elevation, slope, vegetation indices, temperature, humidity, and the De Martonne aridity index, averaged over 100 regression runs.

Training data came from 11,079 samples, combining manual photointerpretation of false-color Landsat mosaics, 59 field-collected GPS points, and 271 points adapted from MapBiomas. The classes followed the IBGE Land Use Classification System and MapBiomas hierarchy across five categories and thirteen level-II classes, from urban areas and aquaculture through croplands, pasture, forest and savanna vegetation, mangrove, herbaceous restinga, and hypersaline tidal flats, to water bodies and non-vegetated areas. Hyperparameter tuning was exhaustive: the number of trees was tested from 100 to 2,000 in ten repetitions per subunit, with out-of-bag error and overall accuracy guiding the choice, and the entire classification pipeline was run 100 times with different random seeds, consolidated by majority vote into a final map for each fragment.

The results were strikingly consistent. Accuracy metrics ranged from 78 to 88 percent and kappa values from 0.75 to 0.85 across the six subunits, with the crystalline portion of the Apodi-Mossoró basin posting the best performance at 0.88 overall accuracy and 0.85 kappa. Forest vegetation and the rivers, lakes, and ocean class achieved precision above 90 percent, and savanna vegetation exceeded 84 percent in crystalline areas. The trouble spots were revealing: salt marshes and herbaceous restinga in the Piranhas-Açu estuarine complex showed inconsistencies, aquaculture proved hard to separate from open water, and croplands were repeatedly confused with pasture, particularly where cassava and cashew grow on sandy soils with sparse cover, or where corn and sugarcane at the start of their cycles reflect light much like grazed land.

An uncertainty index, the normalized variety metric, mapped where the 100 iterations disagreed most. High uncertainty, above 35 percent, clustered in intensively managed croplands such as Serra do Mel and the irrigated fruit-growing hub of the Piranhas-Açu basin, in the heterogeneous urban fabric of Parnamirim and Macaíba, and in the desertification-prone Seridó microregion, where open savanna-steppe vegetation on shallow crystalline soils blurs into pasture. Dense natural vegetation on mountain slopes and massifs, by contrast, showed almost no disagreement, confirming that the model is most confident exactly where the landscape is most coherent.

Perhaps the most consequential finding lies in the variable importance rankings. Across every geomorphological compartment, the most influential predictors were not spectral indices but spatial position, longitude and latitude, altitude, and climatic conditions, namely precipitation and land surface temperature. This makes ecological sense: latitude governs solar input, altitude shapes local temperature and rainfall, and the amount and timing of rain drive the leafing and flowering cycles of seasonally dry vegetation. Surface temperature, meanwhile, separates irrigated fields and dense canopy from exposed soil and degraded pasture. The forest vegetation class, restricted mainly to rainier western and eastern highlands, was especially dependent on these topographic and climatic cues, while water bodies and urban areas remained best distinguished by spectral signatures.

The mapped landscape tells its own sobering story. Native vegetation, whether preserved, degraded, or recovering, still dominates at roughly 58 percent of the territory, about 3 million hectares, with savanna-steppe covering some 2.88 million hectares and forest formations a mere 37,740 hectares concentrated in the wetter ranges. But more than 41 percent of the biome in Rio Grande do Norte now bears anthropogenic uses: pastures span about 1.77 million hectares, croplands around 135,000 hectares, and a growing footprint of wind farms, solar installations, mining, oil extraction, aquaculture, and urban infrastructure reshapes coastal and upland scenery alike. The contrast between the irrigated fruit export economy of the sedimentary Açu-Mossoró hub and the rainfed subsistence farming of the crystalline Sertaneja Depression mirrors the underlying geology, a dichotomy the classifier captured through the sheer weight of longitude and elevation in its decision trees.

When the team compared their geomorphological approach against the traditional method of classifying whole river basins, overall accuracy was statistically similar, but the spatial errors told a different tale. Basin-wide classification inflated croplands in drainage lines, merged urban areas with other uses, and generalized exposed rock and dunes into broad anthropogenic categories, while the terrain-aware segmentation kept those distinctions intact. The authors caution that spatial coordinates can also let a model merely memorize training patterns rather than generalize, a known pitfall of spatially autocorrelated predictors. Still, their conclusion is clear: for seasonally dry tropical forests facing climate projections of rising heat and declining tree cover, mapping systems that understand relief, water, and position, not just reflected light, offer a genuinely better window on one of the world’s most threatened drylands.

Subject of Research: Machine learning land use and vegetation cover mapping in the Caatinga biome using morphometric and climatic variables

Article Title: Integrating morphometric and climatic variables into the mapping of land use and vegetation cover in the Caatinga biome

Article References: Guedes da Silva, A. D., Flor de Souza, S. F., Lucena, R. L., & Reis, J. S. (2026). Integrating morphometric and climatic variables into the mapping of land use and vegetation cover in the Caatinga biome. Environmental Monitoring and Assessment, 198(10), Article 1132. https://doi.org/10.1007/s10661-026-15949-z

Image Credits: AI Generated

DOI: 10.1007/s10661-026-15949-z

Keywords: Caatinga, random forest, land use land cover, Google Earth Engine, remote sensing, semi-arid, Rio Grande do Norte, geomorphology, precipitation, machine learning, dry tropical forest, MapBiomas