Forest Fire is a dataset assembled to build a model that automatically spots wildfires on screen. The labels are simple — three of them. Visible flames mean fire, visible smoke means smoke, and nothing unusual means normal. Aerial drone shots, forest surveillance-camera frames, and on-scene photos are all mixed together, so the shooting conditions vary widely.

The first thing that stands out is the severe skew across the three classes. Smoke accounts for 13,045 images, or 82.8% of the total, while fire is just 1,127. Smoke outnumbers fire by 11.6×. This likely reflects the reality that smoke scenes are far easier to find than flames themselves when collecting wildfire data — but from an AI-training standpoint, it's a signal worth flagging.

Level 1 is a basic fitness test at the pixel level. Overlay every photo in a class and average them, and you get that class's "average face." The sharper the average image, the more consistent its composition and color; the blurrier it is, the more varied the photos inside. Below, each class's actual representative photo (left) sits next to its mean image (right).

Level 2 views the data through the eyes of a general-purpose neural network called Wolfram ImageIdentify Net V2. Each photo turns into 1,280 numbers, with similar photos placed close together and different ones far apart. DataClinic marks how tightly packed each photo sits among its same-class neighbors as a "density." The higher the density, the closer that photo is to its class's typical form.

Look first at the overall density distribution and it's bimodal — two peaks. The lower, smaller peak is made mostly by fire; the higher, larger one by smoke and normal.

Compare the median density per class and the problem sharpens. Fire sits clearly low at 0.11, while smoke is 0.30 and normal 0.33. The majority class, smoke, actually lands slightly below normal. More importantly, the gap between them is just 0.03. Through a 1,280-dimensional general-purpose lens, density alone makes smoke and normal practically impossible to separate.

The same blur shows up in the spatial picture. The PCA below flattens 1,280 dimensions onto a plane. The points don't gather in any one place — they scatter evenly across the frame, so there's hardly any boundary to speak of between classes. Through the general-purpose lens, the wildfire data is still close to a "single undivided fog."

Level 3 is a specialized lens that picks, from the 1,280 dimensions, only those that actually help distinguish the classes — narrowing them down to 17 dimensions. That's roughly 75× compression, but nothing was thrown away blindly: only the axes with a clear discriminating signal were kept. So when the same data is viewed again, the boundary comes out sharper.

The overall density distribution changes first. L2's two peaks vanish, tidying into a single bell-shaped peak bulging in the middle. (The density range grows from L2's to 1.7–9.3 here — about 15× larger — but this isn't because the data improved; it's a mathematical effect of points gathering into a tighter space. So you must not compare L2 and L3 density numbers directly — only within the same level.)

Class separation improves noticeably. The medians spread to fire 3.0, normal 5.2, and smoke 5.9, making the order fire ≪ normal < smoke unmistakable. The smoke-to-normal gap, just 0.03 at L2, widens to about 0.6. The majority class, smoke, rises to the most typical position (highest density), finding its rightful place. Fire has the narrowest box — revealing that while typical flame photos resemble one another, there simply aren't many of them.

The spatial picture is reorganized too. In the L3 PCA, a dense core of points forms at the lower right, while points thin out toward the left. L2's even fog becomes a structure of "one core + scatter." (The PCA points carry no class color. But overlaying the density analysis, the highest-density smoke appears to form this right-hand core, while the lowest-density fire sits in the left-hand scatter.) This is where the work of dimension optimization shows most clearly.

Place each class's density map beside its representative photo, and you can see where the three classes sit in the 17-dimensional space. Fire is isolated in a small upper-left corner, smoke broadly dominates the center and lower right, and normal settles on the periphery — though some of it spills into smoke's territory.

The photos with the highest L3 density — the "most smoke-like" — ranked first and second were both smoke (9.24 and 9.21). But third place (9.14) carried a normal label: normal_457. It trails the top smoke by only 0.1. A photo classified as normal sits in nearly the same spot as the most typical smoke. We looked directly at what it shows.

At the opposite end sits fire. The photo with the lowest L3 density — the one furthest from its own class's type — ranked first (1.68), and it was a fire: fire_392. A typical flame blazes orange across the whole frame, but this photo was the exact opposite.

The two scenes we've seen aren't merely interesting exceptions. They preview how a wildfire AI trained on this data would get things wrong in the field. Two failures, pointing in opposite directions, grow at once from a single dataset.

Forest Fire is not a badly made dataset. The labels are consistent, there are no missing values, and 15,751 images is a respectable scale. But because these are wildfire photos scraped from the real world, they inherit the real world's ambiguity intact. Smoke, fog, and calm scenery blend naturally, and fire always comes with smoke. The 17-dimensional compression pulled apart most of this ambiguity — but the two boundary scenes held out to the end.