IDCubeCloud Documentation
Machine Learning
Create labels, train supervised classifiers, predict classes, and export reusable models.
Classifier
Key Parameters
Practical Guidance
Linear SVM default
No exposed primary parameter
Fast and a strong baseline for spectral data
Logistic Regression
No exposed primary parameter
Linear model with well-behaved probability estimates
k Nearest Neighbours
Number of neighbours k; default 5
No fitting cost but slower prediction; sensitive to feature scaling
Decision Tree
Maximum depth
Easy to interpret but prone to overfitting when deep
Random Forest
Number of estimators; default 100; maximum depth
Robust general-purpose nonlinear model; slower to train
Linear Discriminant Analysis
No exposed primary parameter
Linear discriminant model that assumes approximately Gaussian classes
Gaussian Naive Bayes
No exposed primary parameter
Very fast; assumes conditional independence among bands
Observed Result
Likely Cause
Recommended Check
One class dominates the map
That class contributed many more labelled pixels or covers broader variation
Review sample counts and improve class balance
Speckled or noisy output
Too few samples, mixed pixels, noise, or genuine spectral overlap
Add representative labels and compare with a smoother or more robust classifier
A class is never predicted
Too few examples or spectra nearly identical to another class
Inspect class spectra and add distinct examples
Preview is good but full prediction is poor
Labels represent only a limited part of the scene's variation
Add regions from other locations, illumination conditions, and backgrounds
Limitation
Current Behavior
Recommended Practice
Cube size
Very large cubes may be rejected to protect server memory
Crop the cube or reduce its spatial dimensions before training
Spectral-only labels
Each pixel is classified from its spectrum without texture or neighbourhood context
Interpret isolated noisy pixels cautiously and consider post-processing where appropriate
Class imbalance
Classes with fewer labelled pixels may be under-predicted
Use comparable and representative sample counts
No unknown class
Every pixel is assigned to one known class
Add a background or other class and validate out-of-distribution regions
Cross-scene transfer
A model may transfer poorly across sensors, dates, units, or illumination conditions
Match wavelengths and preprocessing, or retrain with representative data from the new scene
Label leakage
Nearby training and validation pixels may be highly correlated
Use spatially separated regions for an independent assessment
Use Case
Example Classes
Practical Note
Crop type mapping
Corn, soybean, wheat, soil, residue
Label multiple fields and illumination conditions for each crop
Water and land separation
Open water, vegetation, soil, built surface
A threshold mask can seed obvious water labels before refinement
Mineral or soil mapping
Known outcrops, soil types, background
Confirm that the sensor contains bands capable of separating the targets
Damage or stress assessment
Healthy, mildly stressed, severely stressed, background
Ground truth is especially important because stress spectra may overlap
Repeat monitoring
Stable classes from an earlier acquisition
Reuse is most reliable when sensor, wavelengths, scaling, and acquisition conditions match