🐛 Detector de Plagas en Cultivos — Clasificador de Características de Textura en Vivo

For every scanned patch the simulation reads the actual pixel data of the synthetic leaf image and computes three genuine texture statistics: the mean grayscale intensity, the population variance of intensities within the patch (a measure of local roughness), and an edge density obtained by running real Sobel Gx/Gy convolution kernels over every interior pixel and counting how many exceed a gradient-magnitude threshold. None of these numbers are faked or looked up — they are derived directly from the RGBA buffer that also drives what you see on screen.

A reference set of patches is sampled from the same leaf and labelled using ground-truth pest-blotch coverage (only used to build this labelled set, never given to the classifier for unknown patches). Each reference patch's three features are z-score normalised. When a new patch is scanned, its features are normalised the same way, the real Euclidean distance to every reference patch is computed, the k closest are kept, and the patch is flagged damaged if the fraction of damaged neighbours among those k meets the damage-vote-threshold slider — otherwise it is healthy.

Necrotic or chewed leaf tissue tends to have irregular, high-contrast mottling rather than the smooth mottled green of healthy tissue, so the pixel-intensity variance within a damaged patch is higher. The ragged boundary between damaged and healthy tissue also produces far more sharp intensity transitions, which is exactly what a Sobel filter is designed to detect, so edge density spikes near and inside damage.

k sets how many nearest reference patches are consulted for each vote — a small k reacts fast to local texture but is noisier, a large k smooths the decision but can blur small damage spots. The damage-vote-threshold sets what fraction of those k neighbours must be labelled damaged before the patch itself is flagged damaged; lowering it makes the detector more sensitive (more false positives), raising it makes it stricter (more missed damage).

It is a real, if simplified, machine-learning pipeline: genuine feature extraction (variance, Sobel edge density, mean intensity) from real pixel data, a genuine labelled reference set, real Euclidean distance computation, and a real majority-vote k-NN decision rule. It intentionally omits things like data augmentation or a learned metric, but every number driving the classification is computed live from the image, not hand-tagged per patch.