The Dirt Under the Algorithm

The first comprehensive assessment of global soil degradation, published in 1991 under the auspices of UNEP, estimated that approximately 1.9 billion hectares of agricultural land had been degraded to some degree since 1945. Degraded is a deliberately vague word, chosen to paper over disagreements among researchers, but the underlying reality is precise enough: compaction, erosion, salinization, acidification, organic matter depletion, and chemical contamination had reduced the productive capacity of land that humanity depends on for most of its calories.

Thirty-five years later, those processes have continued. The degradation hasn’t reversed. The land itself doesn’t care about policy frameworks or carbon credit markets or precision agriculture software subscriptions. Soil is a living system, about a teaspoon of healthy topsoil contains more microbial organisms than there are humans on Earth, and it responds to management on timescales that don’t line up with quarterly earnings calls or election cycles.

This is the context in which AI soil monitoring is being deployed. Not a clean technological problem, but a deeply tangled biological, economic, and political one.

Traditional soil assessment is painfully slow and expensive. A proper soil test requires physical sampling: someone walks the field, collects cores at specific depths, sends them to a laboratory, waits two to three weeks, and receives a report with nitrogen, phosphorus, potassium, pH, organic matter percentage, and perhaps a dozen other metrics. That report reflects conditions at the moment of sampling. By the time it’s acted on, conditions have changed.

The cost runs between $30 and $150 per sample depending on the depth and analysis requested. A large 1,000-acre operation with heterogeneous soils might reasonably want 50-100 samples to capture spatial variability. At scale, proper soil mapping becomes a significant line item that most farmers can’t afford annually, so they sample every three or five years, making decisions on stale data.

The AI approach to this problem runs on several parallel tracks. Hyperspectral soil sensing, using near-infrared spectroscopy to read soil chemistry from reflected light, has been advancing rapidly. Devices mounted on farm equipment can now estimate organic carbon, nitrogen, pH, and moisture as the tractor moves through the field, generating continuous soil maps at sub-meter resolution during normal field operations. The accuracy is lower than wet chemistry laboratory analysis, but it’s real-time, it’s spatial, and it generates data on every pass rather than every three years.

Satellite-based soil monitoring fills a different gap. Specific spectral bands, particularly in the shortwave infrared, are sensitive to soil organic matter content when the soil surface is bare. This limits satellite approaches to periods between crop cycles, but it allows global monitoring at a scale that field sensors can’t match. The European Space Agency’s Sentinel-2 and Sentinel-3 missions have created a public archive that researchers are mining for soil carbon estimates across enormous areas.

The synthesis, training machine learning models on ground-truth laboratory samples to calibrate both sensor and satellite readings, is where the real capability gains are coming from.

SoilGrids, maintained by the International Soil Reference and Information Centre (ISRIC) in Wageningen, is the closest thing to a global soil intelligence layer that exists. The 2022 version used ensemble machine learning trained on 240,000 soil profiles collected globally to produce maps of 17 soil properties at seven depth increments, globally, at 250-meter resolution. That’s extraordinary compared to what existed a decade ago. It’s also far too coarse for farm-level management decisions.

The commercial players filling the precision gap include Farmers Edge, Pattern Ag, Indigo Agriculture, and several startups that emerged from the intersection of remote sensing research and agtech venture capital. Pattern Ag’s approach, using DNA metagenomic sequencing of soil samples to characterize the microbial community, is genuinely novel. Their argument is that the microbial community is a more sensitive leading indicator of soil health than chemical analysis, because the microbes respond before the chemistry shifts detectably.

Whether this approach will scale outside research settings remains to be demonstrated. Metagenomic sequencing is still expensive per sample, and the reference databases for agricultural soil microbiomes are thin compared to, say, the databases for human gut microbiota. But the principle is sound: a healthy soil food web, bacteria, fungi, nematodes, protozoa, earthworms operating in concert, processes organic matter, fixes nitrogen, suppresses pathogens, and builds structure in ways that no amount of synthetic fertilizer can replicate. If AI can detect when that web is degrading before yield loss becomes visible, intervention becomes possible at a stage where it still matters.

The carbon angle has complicated everything. Beginning around 2020, voluntary carbon markets began paying farmers for increasing soil organic carbon, essentially rewarding practices like cover cropping, reduced tillage, and compost application that build organic matter. The theory is reasonable: soil can store significant amounts of carbon, and shifting billions of acres toward carbon-building management would make a non-trivial dent in atmospheric CO2.

The measurement problem is severe. Carbon markets need to verify that promised sequestration actually happened. Traditional soil sampling is too expensive and too infrequent to support the verification requirements. This is where AI enters as an enabling technology: continuous satellite and sensor monitoring, calibrated by periodic ground truth samples and validated by machine learning models, could theoretically provide the monitoring, reporting, and verification (MRV) infrastructure that carbon markets need.

Companies like Indigo Agriculture, Soil Capital, and the now-defunct Nori built business models on exactly this premise. The execution has been harder than the pitch. Soil carbon measurements have large natural variability, a field can show significant seasonal swings in measured organic carbon from moisture alone, and attributing changes to management practices versus weather and baseline variability requires statistical approaches that are still being refined.

The voluntary carbon market itself collapsed in credibility between 2023 and 2025, following investigations into forest carbon credits that revealed systematic over-crediting. Soil carbon programs survived in somewhat better shape because the measurement methodology is less susceptible to the specific failure modes that plagued forestry credits. But the market uncertainty has slowed investment in exactly the AI monitoring infrastructure that would make soil carbon credits credible.

This is a frustrating feedback loop. Better monitoring technology requires capital investment. Capital investment depends on market demand. Market demand depends on trust in the monitoring technology. The chicken-and-egg problem has been visible for years and hasn’t resolved.

The dimension of this that gets least attention is the deep history encoded in soil itself. Healthy agricultural soil takes roughly 500 years to form one inch of topsoil under natural conditions. American farmers lost an estimated two inches of topsoil in the Dust Bowl years of the 1930s. That loss is not recoverable on any human planning horizon through normal biological processes.

What regenerative agriculture practices, the kinds of practices that AI monitoring systems are designed to encourage and verify, can actually do is slow further loss, stabilize what remains, and gradually build organic matter in the top few inches. That is genuinely valuable. It is not the same as restoration.

The AI systems should be honest about this. The most sophisticated soil monitoring platform on Earth cannot tell you whether the land you’re farming is in the same condition as it was in 1920. It can tell you what happened to that field in the last three years, and it can build predictive models for what specific management choices will do to organic carbon, compaction, and moisture-holding capacity over the next five.

That is useful. More useful, arguably, than any amount of retrospective regret about what industrial agriculture did to soils over the twentieth century. The historical accounting matters for understanding the scale of the problem. The actionable question is whether AI monitoring systems will actually change farmer behavior fast enough to matter.

Early evidence suggests modest positive effects in large commercial operations with the capital and management bandwidth to act on the data. Smaller operations, which comprise the overwhelming majority of global agriculture by farm count if not by acreage, haven’t been reached in any meaningful way. The tools are too expensive, the connectivity requirements are too demanding, and the agronomic interpretation layer, translating sensor readings into specific management decisions, hasn’t been built for the enormous diversity of soils, crops, and farming systems outside North America and Europe.

The dirt doesn’t care about any of this. It just keeps responding to what gets done to it, slowly, on its own timeline, whether or not someone is watching.