Attention-Powered AI Reads Landsat 9 to Map Soil Carbon With New Precision

Soil holds more carbon than the atmosphere and all vegetation combined, yet scientists still struggle to measure it at scale. Digging and lab-analyzing thousands of samples is slow and expensive, which is why researchers have long turned to satellites to infer how much organic carbon lurks in the topsoil of farmland. A new study published in Environmental Monitoring and Assessment pushes that ambition further, introducing an interpretable machine learning framework that squeezes more accuracy out of freely available Landsat 9 imagery than conventional approaches have managed so far.

The research, led by Yi Lv, Chuanhua Zhao, Xinju Li, and Xiao Hu of Shandong Agricultural University in Taian, China, tackles a subtle but important weakness in a popular modeling strategy known as stacking. In a stacked ensemble, several different algorithms, such as random forest, support vector machine, extreme gradient boosting, and a multilayer perceptron, each produce their own estimate of soil organic carbon. A second-level model, called a meta-learner, then blends those predictions into a final answer. The trouble, the authors explain, is that the meta-learners used in most soil studies are linear, meaning they assign each base model a fixed global weight. That works reasonably well, but it hides a crucial reality: the reliability of each algorithm can change dramatically from one sample to the next, depending on where a soil sits along the carbon gradient.

To fix this, the team built AT-FC-Stacking, a framework that combines a sample-wise attention mechanism with a fully connected neural network acting as the meta-learner. Attention, the same idea that powers modern language models, allows the fusion layer to weigh the base models’ predictions differently for every single sample rather than applying one rigid formula to all of them. The fully connected layer then performs an adaptive nonlinear fusion, capturing interactions between the base predictions that a simple weighted average would miss. The result is an ensemble that not only predicts better but also explains itself, because the attention weights reveal exactly how much each base model contributed to any given prediction.

The inputs came from Landsat 9, the newest satellite in the five-decade Landsat record, whose multispectral instrument captures visible, near-infrared, and shortwave-infrared reflectance at thirty-meter resolution. From this imagery the researchers derived a rich set of candidate variables: individual spectral bands, standard vegetation indices such as NDVI, SAVI, and the atmospherically resistant vegetation index, water-related indices, and gray-level co-occurrence texture features that describe the spatial roughness of the landscape. Terrain information from the SRTM digital elevation model added topographic context, since slope and elevation strongly influence where organic matter accumulates.

Not every candidate variable earns its place, however. The team filtered the pool using a two-stage selection procedure that combined Pearson correlation screening with the Boruta algorithm, a wrapper method that compares each feature’s importance against randomized shadow features to confirm it carries genuine signal. This process identified key drivers including Band 2, the blue band, and a texture statistic called sum average, alongside other spectral and topographic covariates. The prominence of the blue band is notable, because soil reflectance in the visible range responds directly to the darkening effect of organic matter, while texture features help the model distinguish bare soil surfaces from vegetated or moist pixels that would otherwise confound the signal.

The performance gains were substantial. Across repeated runs, AT-FC-Stacking achieved a mean coefficient of determination, or R², of 0.594, roughly 10.41 percent higher than the traditional linear stacking benchmark built on linear regression, which averaged 0.538. The attention-based framework also delivered lower root mean square error and lower mean absolute error, meaning its predictions deviated less from laboratory-measured carbon values. Just as importantly, it outperformed every individual base model on its own, confirming that the value comes from intelligent fusion rather than from any single algorithm’s strength.

What sets the study apart, though, is what the attention mechanism revealed about the ensemble’s inner workings. As soil organic carbon values rose along the observed gradient, the contributions of extreme gradient boosting and the multilayer perceptron increased, while the roles of the support vector machine and random forest declined. In other words, different algorithms dominate different parts of the carbon spectrum, and a linear meta-learner is blind to this dynamic. For practitioners, that insight is more than academic curiosity: it explains why a model that excels in one field or region may falter in another, and it points toward ensembles that can be diagnosed and improved rather than treated as black boxes.

The implications extend well beyond modeling elegance. Accurate, repeatable maps of soil organic carbon underpin sustainable agricultural management, national greenhouse-gas inventories, and the verification of carbon sequestration programs that pay farmers for storing carbon in their fields. Measurement, reporting, and verification schemes demanded by international climate agreements require methods that are cheap, transparent, and auditable. Because Landsat 9 data are free and its archive is continuous, an approach built on it can be reproduced anywhere on Earth, and the interpretability layer gives auditors a way to check why the model made each call, something increasingly expected of artificial intelligence deployed in environmental policy.

The work also fits into a broader shift in digital soil mapping. Over the past decade, researchers have moved from simple regressions on a few spectral indices toward multi-source data fusion, deep learning, and ensemble methods, drawing on Sentinel-1 and Sentinel-2 radar and optical data, hyperspectral sensors, and environmental covariates. Yet many of these sophisticated models sacrifice transparency for accuracy, and studies have repeatedly shown that sample size, soil texture, surface moisture, and residue cover can all destabilize predictions. By making the fusion step itself interpretable, the Shandong team offers a template for reconciling the two goals, showing that accuracy and accountability need not be traded off.

Limitations remain, as the authors acknowledge in framing their contribution. The framework was developed and evaluated with field data from a specific region, and its attention patterns may shift in landscapes with different land uses, climates, or soil types. The data availability statement notes that no new datasets were generated or analyzed beyond the study itself, so independent validation on other continents will be the next test. Still, the message is clear: the next generation of soil carbon maps may come not from a single heroic algorithm but from committees of models that know when to defer to one another, observed from seven hundred kilometers above the ground by a satellite that has been quietly watching Earth’s soils for half a century.

Subject of Research: Predicting soil organic carbon content from Landsat 9 satellite imagery using interpretable stacking ensemble machine learning

Article Title: The capability of SOC content prediction by employing two stacking ensemble learning models based on Landsat 9 imagery

Article References: Lv, Y., Zhao, C., Li, X., & Hu, X. (2026). The capability of SOC content prediction by employing two stacking ensemble learning models based on Landsat 9 imagery. Environmental Monitoring and Assessment, 198(10), Article 1085. https://doi.org/10.1007/s10661-026-15923-9

Image Credits: AI Generated

DOI: 10.1007/s10661-026-15923-9

Keywords: soil organic carbon, Landsat 9, stacking ensemble learning, attention mechanism, machine learning, remote sensing, digital soil mapping, Boruta feature selection, XGBoost, random forest, interpretability, agriculture