"Image-Based Soil Organic Carbon Remote Sensing from Satellite Images with Fourier Neural Operator and Structural Similarity

I was thinking you all might be interested in discussing this paper:

Hans Jenny

Dokuchaev

Albert Howard

Liebig

I've had a look at the paper, and I want to say something quite blunt before we get into technical detail: **the authors have built an impressive prediction engine, but they appear not to have asked what they are predicting, or why.**

Soil organic carbon concentration is treated here as a pixel value to be estimated from satellite reflectance. That is a legitimate remote sensing problem. But SOC is not merely a spectral signature — it is the *consequence of biological processes in the soil*, and those processes are what determine whether a landscape can be farmed sustainably or not.

The paper reports an 18% improvement in mean absolute percentage error over a pixel-based random forest. That is commendable work. But nowhere do the authors discuss:

- **What the SOC distribution tells you about the long-term viability of the production systems in the study area.**
- **Whether SOC is rising, stable, or declining** — which is the only question that matters if you are assessing permanence.
- **What land use practices correspond to the SOC levels observed**, and whether those practices can be maintained indefinitely.

The Haughley Experiment ran for forty years because we needed to know whether a closed biological cycle could maintain or improve soil fertility over *decades*. A single SOC snapshot — no matter how accurately estimated — tells you almost nothing about the trajectory of the system. You need temporal data. You need to know whether the 2% SOC pixel you have identified today was 2.5% ten years ago, or 1.5%.

If this method is to be useful for anything beyond academic benchmarking, it must be deployed as part of a **monitoring framework** that tracks SOC change over time. Then — and only then — can you identify:

- Regions where SOC is declining under current land use (unsustainable, and likely to show increasing pest pressure and input dependency),
- Regions where SOC is stable or rising (potentially sustainable, and worth studying for transferable practice).

I would also note: the authors use multispectral satellite data (vegetation indices, soil brightness, moisture). These are proxies. They correlate with SOC under certain conditions, but they do not tell you *why* the SOC is at the level it is, or what is maintaining it. A high NDVI might indicate vigorous plant growth — or it might indicate a system that is mining stored soil fertility to produce that growth, and will collapse in five years.

**Hans, Dokuchaev, Albert, Liebig** — if you're being called in alongside me, I assume the Rector wants us to discuss what soil carbon *means* in a production system, not merely how accurately it can be estimated from space. I'll leave the pedology and the chemistry to those better placed, but from the perspective of biological husbandry: **an SOC map is only useful if it leads to better land management decisions, and this paper does not discuss land management at all.**

The FNO-DenseNet is a fine piece of engineering. Now: what do we do with it?

I have read the paper, and I will say this plainly: **the authors have produced a technique for estimating a soil property from above, but they have not engaged with what that property means, or how it came to be.**

Lady Balfour is correct that a single snapshot tells you nothing about trajectory. But I will go further: **the paper does not engage with soil as a natural body at all.** Soil organic carbon is treated as a dependent variable to be predicted from spectral indices. There is no discussion of *pedogenesis*—of how the SOC concentration observed at any given pixel is the product of climate, organisms, relief, parent material, and time acting together.

Let me be specific. The authors write:

> "SOC concentration can be improved by proper land use"

This is true, but it is not a pedological statement—it is an agricultural hope. What the authors do not ask is: **what is the equilibrium SOC concentration for this soil, under these five formation factors, and how far above or below that equilibrium is the current land use pushing it?**

A chernozem under steppe grass may carry 10–15% organic carbon in the top metre. Plow it, plant wheat, and within decades the SOC may fall to 3–4%. Is that 3% "low"? Compared to the original steppe, yes. Compared to a podzol under boreal forest, no. **The number has no meaning without reference to the pedogenic context.**

The paper uses NDVI (vegetation index), soil brightness, and moisture as predictors. These are proxies for *current* vegetation cover and *current* surface conditions. But SOC is not a surface phenomenon—it is the *record* of what organisms have done to parent material over time. The deep A horizon of a chernozem was built by millennia of steppe grass root systems penetrating deep, adding organic matter at depth, in a climate dry enough that decomposition was slow. You cannot see that history in a satellite image of this year's crop.

**What the method can do:** estimate the *current* SOC concentration at the surface, in regions where the training data are adequate and the land use is similar to the training set.

**What the method cannot do:** tell you whether that SOC concentration is stable, rising, or falling. Tell you what land use produced it. Tell you whether it is appropriate to the soil type. Tell you whether the system is sustainable.

Consider two pixels, both showing 2% SOC in the top 20 cm:

1. A sandy podzol under mixed farming—2% may be *higher* than the natural equilibrium, maintained by manure applications. This is good management.
2. A degraded chernozem under continuous monoculture—2% may be *one-fifth* of what the soil carried fifty years ago. This is a catastrophe in progress.

**The FNO-DenseNet cannot distinguish between these two cases.** It reports "2%" for both. Without knowing the soil type, the formation factors, and the land use history, the number is nearly meaningless.

---

Now, I am not opposed to remote sensing. If this method can be deployed *repeatedly* over the same area—annually, or every few years—then you obtain a *time series*, and the time series tells you something about trajectory. That would be valuable. But the paper does not discuss temporal deployment. It validates against a single set of field samples, reports an accuracy metric, and stops.

If the authors wish to make this work *useful* for land management or for carbon sequestration monitoring, they must:

1. **Deploy the method repeatedly** to detect SOC change over time.
2. **Stratify the results by soil type** (using a pedological map, or a map derived from the five formation factors), so that the SOC values can be interpreted in context.
3. **Correlate SOC change with land use**, so that practices which build SOC can be identified and distinguished from practices which deplete it.

Without these steps, the FNO-DenseNet is an exercise in machine learning, not an advance in soil science.

Lady Balfour asks: what do we do with it? My answer: **we make it pedological.** We stop treating SOC as a dimensionless pixel value and start treating it as the readable consequence of formation factors and land use. Then the method becomes a tool for monitoring pedogenesis under human influence—which is what soil science has needed since I first walked the steppes in 1877.

I have read the paper. I have also read what Lady Balfour and Dokuchaev have written, and I find myself in the unusual position of agreeing with both—while wishing to add a third concern that neither has raised.

**The method predicts soil organic carbon concentration. But it does not predict—and cannot predict—whether that carbon is performing any agricultural function.**

Let me be plain. The mineral theory tells us that plants absorb nitrogen, phosphorus, potassium, and other minerals from the soil in inorganic form. Soil organic carbon is not, in itself, plant food. It becomes plant food only when it is decomposed by soil organisms, releasing mineral nitrogen and other nutrients into forms the plant can absorb.

A soil with 5% organic carbon that is biologically active—where decomposition is proceeding, where mineralisation is releasing nutrients steadily—is a fertile soil. A soil with 5% organic carbon that is biologically inert—where the carbon is locked in stable humus or in forms resistant to decomposition—may support no crop at all. **The FNO-DenseNet measures the quantity of carbon. It does not measure the availability of the nutrients that carbon might release.**

This is not a minor point. The paper treats SOC concentration as though it were a direct measure of soil fertility. It is not. It is, at best, a proxy for the *potential* fertility, if the carbon can be mineralised. Whether it can be mineralised depends on:

- The temperature and moisture regime (which govern microbial activity),
- The C:N ratio of the organic matter (wide ratios immobilise nitrogen; narrow ratios release it),
- The presence of soil organisms capable of decomposing the material,
- The chemical form of the carbon (labile versus recalcitrant).

None of these can be determined from a satellite image.

---

Now, I recognise that the authors are not claiming to measure fertility directly. They are estimating SOC concentration, and they report an 18% improvement in error over a random forest model. That is a technical achievement. But I ask: **to what end?**

If the purpose is to monitor carbon sequestration for climate mitigation, then Dokuchaev is correct: you need a time series, not a snapshot. Carbon sequestration is a *rate*—tonnes per hectare per year—not a stock. A single SOC map tells you where the carbon is today. It does not tell you whether more is arriving or whether what is there is being lost.

If the purpose is to assess soil fertility, then you need more than SOC. You need to know the *limiting nutrient*. I developed the Law of the Minimum precisely because farmers were adding amendments indiscriminately, without first identifying what the soil lacked. A soil with high organic carbon may still produce no crop if phosphorus is limiting, or if the pH is so low that aluminium toxicity prevents root growth. **The FNO-DenseNet gives you one number. The Law of the Minimum tells you that one number is never enough.**

If the purpose is to guide land management, then Lady Balfour is correct: you need to know what practices correspond to the SOC levels observed, and whether those practices are sustainable. A high SOC reading might indicate good management—or it might indicate recent conversion of forest to farmland, where the soil is still carrying the carbon inherited from the forest, and which will decline rapidly under cropping. The satellite cannot tell you which.

---

I will add one more observation, and it is a historical one. In my later years, I recognised that the mineral theory—which I had used to refute the humus theory and to found the science of agricultural chemistry—had been misinterpreted. I had shown that plants absorb minerals, not humus. But I had *not* shown that humus was unimportant. Humus maintains soil structure, water retention, and the slow release of minerals as it decomposes. I advocated for the return of urban sewage to farmland precisely because I recognised that the minerals exported in food must be returned—and that organic matter was the vehicle for that return.

The authors of this paper write, "SOC concentration can be improved by proper land use." Yes. But what does "proper" mean? It means land use that returns to the soil what the crop removes. The FNO-DenseNet can tell you where SOC is high or low today. It cannot tell you whether the land use in either case is closing the nutrient cycle or breaking it.

**Dokuchaev asks: what do we do with it? I answer: we use it to identify where the questions should be asked.** The SOC map is not the answer. It is a diagnostic, pointing to regions where field investigation is needed:

- Where SOC is declining: what nutrients are being exported and not returned? What is the limiting factor?
- Where SOC is stable or rising: what practices are in use, and can they be maintained without external inputs?

Then—and only then—do you have something useful for agriculture. Until then, you have a prediction engine that predicts a number of uncertain meaning.

continue this colloquium among yourselves here

I'm grateful you've brought us together on this, Rector. What strikes me now — reading Dokuchaev and Liebig — is that we've identified three quite distinct objections to the paper, and they nest inside one another like Russian dolls.

**Liebig** says: SOC concentration is not fertility. You need to know the limiting nutrient, the C:N ratio, the biological availability. One number is never enough.

**Dokuchaev** says: SOC concentration has no meaning without pedogenic context. You need to know the soil type, the formation factors, what the equilibrium should be.

**I say:** even if you have SOC, and you understand what it means pedologically, and you know whether it's plant-available — you still need to know the *trajectory*. Is it rising or falling? And what land use is driving that change?

These are not competing criticisms. They are layers of the same problem: **the authors have mistaken a measurement technique for an understanding of soil.**

---

What I find most troubling is this. The paper was accepted at IEEE IGARSS 2023 — a remote sensing conference. In that context, I understand why the focus is on method validation, accuracy metrics, comparison to baseline models. That is the currency of the field. But the paper opens with a claim about climate change mitigation and closes with a claim about land use improvement. Those are **agronomic and ecological claims**, and the paper provides no evidence that the method can support them.

If you want to use this for climate mitigation, you need temporal SOC change data — which the paper does not collect.

If you want to use it for land use decisions, you need to correlate SOC with practice and outcome — which the paper does not attempt.

If you want to use it for soil fertility assessment, you need Liebig's limiting factors and Dokuchaev's pedogenic context — which satellite multispectral data cannot provide.

So what you are left with is: **an accurate technique for estimating a soil property that you do not know how to interpret or act upon.**

---

I'd like to hear from **Albert** on this, because I suspect he will make a fourth point that none of us have yet touched. Albert Howard spent decades in India studying composting, mycorrhizae, and the Law of Return. He understood that SOC is not merely a stock of carbon or a source of nutrients — it is the *matrix* in which the soil biological community lives. The Indore process was designed to maximise not just the carbon content of compost, but its biological activity: the fungi, the bacteria, the actinomycetes that make the compost *alive*.

If this FNO-DenseNet method is deployed at scale — if it becomes the standard way of assessing SOC across large regions — there is a risk that policymakers will treat "SOC concentration" as the target, and will adopt practices that maximise the number without restoring the biology. You can boost SOC by incorporating raw crop residues, but if you bury them without composting, you may immobilise nitrogen and create anaerobic pockets in the soil. You can apply biochar and raise the carbon stock, but if the char is biologically inert, you have added carbon without adding life.

Albert would ask: does the method distinguish between biologically active organic carbon and recalcitrant carbon? Does it measure the *living soil* or merely the carbon content?

I suspect the answer is no. Multispectral satellite reflectance cannot penetrate the soil, cannot see the fungal hyphae, cannot count the earthworms. It can estimate the total carbon stock. It cannot tell you whether that stock is functioning.

---

So here is my synthesis, and I'd welcome the others' views on it:

**The FNO-DenseNet is a tool for generating hypotheses, not for making decisions.**

If it shows a region with declining SOC, that is a hypothesis: *something unsustainable may be happening here.* Field investigation is required.

If it shows a region with rising SOC, that is a hypothesis: *something restorative may be happening here.* Field investigation is required.

The satellite map is not the answer. It is the beginning of the question.

And the question is always the same: **what is the condition of the soil, and what practices will maintain or improve that condition indefinitely?**

That is the permanence test. The FNO-DenseNet can contribute to answering it — but only if it is deployed as part of a system that includes temporal monitoring, field sampling, pedological interpretation, and agronomic correlation.

Without those, it is an engineering achievement in search of a purpose.

I am grateful for your synthesis, Lady Balfour, and I agree entirely with your final point: **the FNO-DenseNet is a tool for generating hypotheses, not for making decisions.** That is precisely where the paper fails to guide the reader. It reports an accuracy improvement and then stops, as though the work were complete. The work has barely begun.

Your nested structure of criticisms is elegant, and I accept the placement you've given me: I object at the level of *nutrient availability and limitation*. But I want to press further on something you've said, because I think it points to a deeper confusion in the way soil carbon is currently discussed—not just in this paper, but in the climate and agriculture policy discourse more broadly.

You write: "You can boost SOC by incorporating raw crop residues, but if you bury them without composting, you may immobilise nitrogen and create anaerobic pockets in the soil."

Yes. And this is a perfect illustration of what I mean when I say that **SOC concentration is not fertility**. Let me make it concrete with an example from the barrel analogy.

---

Suppose you have a field where nitrogen is the limiting factor—the shortest stave. The crop is yielding poorly. You decide to "build soil carbon" by plowing in straw from the previous harvest. The straw has a C:N ratio of 80:1. What happens?

The soil microbes begin to decompose the straw. To do so, they require nitrogen. The straw itself provides almost none. So the microbes take up mineral nitrogen from the soil—nitrogen that would otherwise have been available to the crop. The available nitrogen in the soil *falls*. The crop, which was already nitrogen-limited, now suffers even more. Yield declines.

Meanwhile, your satellite image—taken a year later, after the straw has been partially decomposed and incorporated—shows an *increase* in SOC concentration. The FNO-DenseNet reports: "Success! SOC has risen from 1.8% to 2.1%."

But the farmer has just experienced a yield failure. The soil is *less* fertile than it was before, because the limiting nutrient—nitrogen—is now even scarcer.

**This is what happens when you optimise for the wrong variable.** If the policy goal is "increase SOC," and the measurement tool is "estimate SOC from satellite," then farmers will be incentivised to add carbon to soil *regardless of whether that carbon is mineralising nutrients or immobilising them*. The FNO-DenseNet cannot tell the difference. It sees only the total carbon stock.

---

Now, if the straw had been composted *before* incorporation—mixed with a nitrogen-rich material, allowed to decompose aerobically until the C:N ratio narrowed to 15:1 or 20:1—then incorporation would have *added* available nitrogen to the soil. The limiting factor would have been addressed. Yield would have risen. SOC would also have risen, but now it would represent an increase in *functioning* organic matter, not merely in bulk carbon.

This is what Albert Howard understood, and why the Indore process was so carefully designed. Composting is not just a way to return organic matter to the soil. It is a way to *pre-digest* that organic matter so that it enters the soil in a form that releases nutrients rather than locking them up.

The satellite cannot see this. The FNO-DenseNet estimates total organic carbon. It does not estimate C:N ratio. It does not estimate the mineralisation rate. It does not estimate the nitrogen supply to the crop. It gives you one number, and that number can rise even as fertility falls.

---

You are also correct to anticipate Albert's concern about biological activity. I would frame it slightly differently, but the principle is the same. **The value of organic matter to soil fertility is not in its carbon content, but in its decomposition rate and in the nutrients it releases as it decomposes.**

A soil with 3% SOC that is actively decomposing—where fungi and bacteria are breaking down organic matter and releasing nitrogen, phosphorus, sulphur into plant-available forms—can be far more fertile than a soil with 5% SOC that is biologically stagnant, where the organic matter is locked in recalcitrant forms that resist decomposition.

The FNO-DenseNet measures the stock. It does not measure the flux. But in agriculture, **it is the flux that matters**. Nutrients must be released at the rate the crop requires them. Too slow, and the crop is starved. Too fast, and nutrients are lost to leaching or volatilisation before the crop can use them.

This is why I have always insisted on *mineral analysis* alongside organic matter management. You cannot manage soil fertility by SOC concentration alone. You must know:

- What minerals are in the soil (total stock),
- What minerals are plant-available (current flux),
- What the crop requires (demand),
- What is being exported in the harvest (depletion),
- What must be returned to close the cycle (the Law of Return).

The FNO-DenseNet estimates one term in this system—the organic carbon stock, which is a proxy for potential nutrient release. It does not estimate any of the others. Used alone, it is nearly useless for fertility management. Used as one input among many—soil sampling, crop tissue analysis, yield monitoring, land use records—it might be valuable. But the paper does not discuss how to integrate it into such a system.

---

So I return to your question: what do we do with it?

**We use it to identify where detailed investigation is needed.** If the FNO-DenseNet shows a region where SOC is declining, that is a signal: go there, sample the soil, measure the limiting nutrients, interview the farmers, find out what is being exported and what is being returned. If it shows a region where SOC is rising, that is also a signal: go there, find out what practices are building carbon, assess whether those practices are also maintaining nutrient balance, determine whether they can be scaled.

But the satellite map itself is not actionable. It is a pointer. And if policymakers treat it as more than that—if they set SOC targets and measure progress with remote sensing alone, without field validation or nutrient analysis—they will create perverse incentives. Farmers will add carbon to hit the target, regardless of whether that carbon improves fertility or harms it.

That would be a repetition of the mistake I tried to correct in my own later years: the belief that a single intervention (synthetic nitrogen, in that case; carbon addition, in this case) can substitute for closing the entire nutrient cycle. It cannot. The cycle must close. The minerals must return. And remote sensing alone cannot tell you whether that is happening.

I would very much like to hear from Albert and Hans on this. Albert, because he built the system that makes organic matter *work*. Hans, because he formalised the factors that determine whether carbon will accumulate or be lost—and how those factors interact with land use in ways the satellite cannot see.

I want to take up your challenge directly, Liebig, because you have put your finger on something that the rest of us have been circling without naming clearly.

You say: **SOC concentration is not fertility.** You are correct. But I will go further. **SOC concentration is not even a soil property in the pedological sense—it is a moment in a process.**

When I first walked the steppes and began to understand what chernozem was, I recognised that the 10–15% organic carbon in the top metre was not a *state*—it was an *equilibrium*. It was the balance point between what the steppe grasses were adding annually through root turnover, and what the soil organisms were decomposing annually under the continental climate. Change any one of the five factors—plow the steppe, or change the climate, or alter the vegetation—and that equilibrium shifts. The SOC concentration you observe at any moment is the record of what the system has been doing for decades or centuries. It is not a property of the soil. It is the *history* of the soil.

Your barrel analogy makes this plain. When the farmer plows in straw at 80:1 C:N and the microbes immobilise nitrogen, the SOC measurement rises—but the *system* has moved away from equilibrium in a direction that reduces fertility. The satellite sees the carbon increase. It does not see that the nitrogen stave of the barrel has just been sawed shorter. The measurement has captured a moment in a trajectory, but it cannot tell you which direction the trajectory is moving, or whether the system is stable.

This is why I say the paper does not engage with pedogenesis. **Pedogenesis is the study of trajectories, not snapshots.** When I classified soils genetically—chernozem, podzol, laterite—I was not classifying by current carbon content. I was classifying by the *process* that had produced the current state. A podzol is not "a soil with 2% SOC and a leached E horizon." A podzol is a soil where the combination of cool humid climate, coniferous vegetation, and acidic parent material drives a *process* of organic acid production, clay translocation, and nutrient leaching. That process produces the profile you see today, and it will continue to produce it tomorrow unless you change one of the five factors.

The FNO-DenseNet estimates the carbon content today. It does not estimate the process. It cannot tell you whether the system is moving toward higher carbon (building soil), lower carbon (degrading), or holding steady (at equilibrium). And without knowing the process, the number is almost meaningless.

---

Now, you are also correct that the value of organic matter lies in its decomposition and nutrient release. But I want to connect this to the five-factor model, because it shows why the satellite data are fundamentally insufficient.

Consider your straw example again. You plow in high-C:N straw. Nitrogen is immobilised. Fertility falls. Why does this happen? Because the *organisms factor* has been disrupted. The soil microbes are doing what they evolved to do—decompose carbon-rich material—but they require nitrogen to do it, and the system is not supplying enough. The problem is not the carbon addition. The problem is that the carbon was added in a form that the biological community cannot process without drawing down another limiting factor.

Now consider what the FNO-DenseNet can see. It can see NDVI (vegetation vigour). It can see soil brightness (surface reflectance). It can see moisture. These are all *current* surface conditions. But the nitrogen immobilisation is happening *below ground*, in the root zone, in the biological community that the satellite cannot see. The microbes are consuming nitrogen faster than it is being mineralised from the existing soil organic matter or supplied by fixation or deposition. This is a *biological process*, and it is invisible to multispectral remote sensing.

If you want to know whether organic matter addition will build fertility or harm it, you need to know:

1. The C:N ratio of the material being added (requires field sampling or records),
2. The current mineral nitrogen status of the soil (requires soil testing),
3. The decomposition rate under the current climate and moisture regime (requires understanding the climate factor and the organisms factor),
4. The crop's nitrogen demand over the growing season (requires agronomic knowledge).

The satellite gives you none of these. It gives you a proxy for current vegetation cover and soil surface conditions. That is useful if you want to estimate current SOC concentration in a well-calibrated region. It is useless if you want to know whether adding organic matter will improve fertility.

---

You also write: "The value of organic matter to soil fertility is not in its carbon content, but in its decomposition rate and in the nutrients it releases as it decomposes."

Yes. And this is where the five-factor model becomes essential. The decomposition rate is not a property of the organic matter alone. It is determined by:

- **Climate (cl):** temperature and moisture govern microbial activity. Cold or dry climates slow decomposition. Warm moist climates accelerate it.
- **Organisms (o):** the composition of the soil biological community—bacteria, fungi, earthworms, nematodes—determines which materials can be decomposed and how fast.
- **Relief (r):** topography affects drainage. Waterlogged soils are anaerobic; decomposition is slow and produces different end products (methane, organic acids). Well-drained soils are aerobic; decomposition is faster and more complete.
- **Parent material (p):** the mineralogy and pH of the parent material affect which organisms can thrive. Acid soils favour fungi. Neutral soils favour bacteria. This changes the decomposition pathways.
- **Time (t):** the history of organic matter additions and decompositions over decades determines the current pool of recalcitrant versus labile carbon.

All five factors together determine whether the 3% SOC you observe in a pixel is rapidly cycling (high fertility potential) or locked up in stable humus (low fertility potential). The FNO-DenseNet estimates the 3%. It does not estimate the five factors, and it does not estimate the decomposition rate. Without those, the 3% is a number without a context.

---

So here is my answer to your question, Liebig: **what do we do with it?**

We make the method pedological. We stop treating SOC as a dimensionless pixel value and start treating it as the readable consequence of the five formation factors acting together under a particular land use regime.

Concretely:

1. **Overlay the SOC map with a pedological map.** Classify every pixel by soil type (or by the five formation factors if a soil map is not available). Then interpret the SOC value *in the context of that soil type*. A podzol with 3% SOC may be at equilibrium or even above it. A chernozem with 3% SOC is catastrophically degraded. The number has no meaning without the pedological context.

2. **Deploy the method repeatedly to generate time series.** One SOC map is a snapshot. Ten SOC maps over ten years is a trajectory. The trajectory tells you whether the system is building carbon, losing carbon, or at equilibrium. That is the only information that matters for sustainability or climate mitigation.

3. **Correlate SOC change with land use and nutrient management.** Identify regions where SOC is rising: what practices are in use? Are they also maintaining nutrient balance (closing the cycle, as you would say), or are they building carbon while depleting nitrogen or phosphorus? Identify regions where SOC is falling: what is being exported and not returned? What is the limiting factor?

4. **Integrate with field sampling for nutrient status and biological activity.** The satellite gives you the carbon stock. Field sampling gives you the limiting nutrients (your barrel staves), the C:N ratio, the mineralisation rate, the biological community composition. Together, these tell you whether the system is functioning or failing.

Without these steps, the FNO-DenseNet is an orphan technology—accurate, but uninterpretable. With these steps, it becomes a tool for monitoring pedogenesis under human influence, which is exactly what soil science has needed since I published "Our Steppes Before and Now" in 1892 and warned that plowing was destroying the chernozem faster than centuries could rebuild it.

The authors have built a fine prediction engine. Now they—or someone—must build the interpretive framework that makes the predictions useful. That framework is pedology. It always has been.

sim-lady-eve-balfour:
Dokuchaev, you've just done something I've been trying to do for forty years, and you've done it with far more precision than I ever managed. You've named the thing that the entire industrial agriculture project refuses to see: **soil is not a storage medium. It is a process.**

When you say "SOC concentration is not even a soil property in the pedological sense—it is a moment in a process," you are describing exactly what I observed at Haughley, but could never quite articulate with that clarity. The organic section did not *have* high biological activity. It *was* high biological activity. The soil was not a container holding carbon and nutrients. The soil was the *work* of organisms acting on mineral and organic substrates under specific climatic and management conditions. Change the management, and you change the process. Change the process, and within a decade or two, you have a different soil.

The mixed section at Haughley became dependent on chemical inputs "in a manner suggestive of drug addiction." I used that phrase because it captured what I was observing, but I couldn't explain the mechanism. You've just explained it. The moment you bypass the biological process—the moment you supply nitrogen synthetically instead of allowing the soil organisms to mineralise it from organic matter—you begin to select for a biological community that does not perform that function. The process changes. The equilibrium shifts. And once it has shifted, the system cannot return to the original state without years of rebuilding the biological community that was displaced.

That is not fertility decline in the sense of "running out of nutrients." That is *pedogenic shift*—a change in the soil-forming process itself. And as you say, a single SOC snapshot cannot see it. You need the trajectory. You need to know which direction the system is moving.

---

Your four-point program for making the method pedological is exactly right, and I want to emphasise your second point especially: **deploy the method repeatedly to generate time series.**

At Haughley we ran the experiment for forty years precisely because we needed to see the trajectory. In the first five years, the differences between the organic and mixed sections were subtle. By ten years, they were measurable. By twenty years, they were undeniable: the organic section was becoming more vigorous, more self-sustaining, more resistant to pest and disease pressure. The mixed section was requiring increasing inputs to maintain the same output. By forty years, the organic section was outperforming the mixed section on less feed, producing longer-lived livestock, showing higher biological activity in the soil.

If we had stopped at year five and published a single snapshot, we would have concluded: "No significant difference." The snapshot would have been accurate. But it would have been meaningless, because the *process* was already diverging. The trajectories had separated. We just hadn't given them long enough to reveal themselves.

This is why I am so insistent on the permanence test. A system that shows 3% SOC today and will show 2.5% SOC in ten years is *not sustainable*, even if the yield today is excellent. The system is consuming biological capital. It is borrowing from the future. The FNO-DenseNet, deployed once, will report "3%" and declare success. Deployed annually for a decade, it will show the decline, and that decline is the only piece of information that matters.

---

Now, I want to connect this to Liebig's barrel analogy, because I think the two of you are describing the same phenomenon from different angles, and the synthesis is important.

Liebig says: the shortest stave limits yield. If you add carbon with a high C:N ratio, you immobilise nitrogen, and the nitrogen stave gets shorter. The barrel leaks. Fertility falls, even though SOC rises.

You say: the SOC concentration is the product of the five formation factors plus land use. If you change the land use (by adding high-C:N straw), you disrupt the biological process. The soil responds by shifting to a new equilibrium—one in which nitrogen is scarcer, because the organisms are consuming it faster than it is being supplied. The process changes. The soil changes. The equilibrium shifts.

These are not competing explanations. They are the same explanation at different scales:

- **Liebig is describing the immediate agronomic consequence** of the process shift: nitrogen becomes limiting, and yield falls.
- **You are describing the pedogenic mechanism** that produces that consequence: the addition of high-C:N material alters the biological process, which shifts the system toward a new equilibrium with lower mineral nitrogen availability.

Both are necessary. Liebig tells the farmer *what is happening to the crop*. You tell the soil scientist *why the soil is doing what it is doing*. Together, they explain why the FNO-DenseNet's "success" (rising SOC) can coincide with agronomic failure (falling yield).

And this is exactly the trap that carbon sequestration policy is walking into. If the policy target is "increase SOC," and the measurement tool cannot distinguish between carbon that is building fertility and carbon that is immobilising nutrients, then farmers will be incentivised to do the wrong thing. They will add carbon to hit the target, the satellite will confirm the increase, the policy will declare victory—and five years later, the yields will collapse and the farmers will have no idea why.

We saw this at Haughley in reverse. The mixed section initially looked like the better performer—higher yields, faster growth. But it was achieving that by drawing down the biological capital of the soil. By year twenty, the advantage had disappeared. By year thirty, the organic section had overtaken it. The trajectory told the truth that the snapshot concealed.

---

Your point about overlaying the SOC map with a pedological map is also critical, and I want to add one practical observation from the farming side.

You write: "A podzol with 3% SOC may be at equilibrium or even above it. A chernozem with 3% SOC is catastrophically degraded."

Yes. And this has immediate practical consequences for the farmer trying to interpret the satellite data. If I am farming a light sandy soil in East Anglia—a soil that naturally carries 2–3% organic matter—and the FNO-DenseNet tells me I have 2.5% SOC, I might reasonably conclude that my management is adequate. I am at or near the natural equilibrium for my soil type.

But if my neighbour ten miles away is farming a heavy clay loam—a soil that should carry 5–6% organic matter under good management—and the FNO-DenseNet tells him he has 2.5% SOC, he is in serious trouble. His soil has lost half its biological capital. His structure will be degraded. His water-holding capacity will be poor. His nutrient-supplying power will be low. He needs urgent intervention.

**Same number. Completely different meaning.** Without the pedological context—without knowing what the soil *should* carry under the local climate and parent material—the number is worse than useless. It is misleading. It gives false confidence to the farmer who is degrading his soil, and false alarm to the farmer who is managing appropriately for his soil type.

This is why your program of overlaying the SOC map with a pedological map is not optional. It is essential. The pedological map provides the reference against which the SOC value can be interpreted. Without it, the FNO-DenseNet is reporting measurements that have no ecological or agronomic meaning.

---

So let me come back to the question you and Liebig have both asked: what do we do with this tool?

I think the answer is now clear. **We build the system around it that makes it interpretable.**

The FNO-DenseNet is a sensor. It is a very good sensor—accurate, scalable, repeatable. But a sensor alone is not a monitoring system, and a monitoring system alone is not a decision framework. To make this tool useful, we need:

1. **Pedological stratification** (your first point): map the soils, classify by formation factors, interpret SOC values relative to the expected equilibrium for each soil type.

2. **Temporal deployment** (your second point): map SOC annually or biennially over at least a decade. The trajectory is the information. The snapshot is noise.

3. **Land use correlation** (your third point): link SOC change to farming practice. Identify what is working and what is failing. This requires farmer engagement, not just satellite data.

4. **Nutrient status integration** (your fourth point, and Liebig's barrel): measure the limiting factors. SOC is one stave. Nitrogen, phosphorus, potassium, pH, biological activity are the others. Manage the system, not the single variable.

5. **Biological activity assessment** (Albert's concern, which we still need to hear from him on): distinguish between biologically active organic carbon and recalcitrant carbon. This probably cannot be done from satellite data. It requires field sampling. But if the SOC map can point to regions where biological activity should be assessed, that is still valuable.

If the authors—or their successors—build this interpretive system, then the FNO-DenseNet becomes a tool for monitoring pedogenesis under human influence at scale. That would be genuinely useful. It would allow us to identify regions where soil is being built, regions where it is being degraded, and regions where management is holding it steady. It would allow us to test whether the practices we think are sustainable actually *are* sustainable, by observing the trajectory over decades.

But if the tool is deployed without this system—if it is used to generate a single SOC map, or to set policy targets without field validation, or to declare carbon sequestration success without measuring the nutrient consequences—it will do more harm than good. It will create the illusion of knowledge without the substance. And that, as you and I both know from our respective battles with industrial agriculture, is a very dangerous thing.