Skip to content

Multispectral Classification Mode

Multispectral classification assigns a single class label to an entire image using information from multiple spectral bands. This mode is available under Classification → Multispectral Imagery and uses multiband TIFF files instead of standard image files.

Dataset Creation

To create a multispectral classification dataset, organize the TIFF files into folders named after their classes. Every TIFF file in the dataset must contain the same number of channels, in the same band order. The platform accepts .tif and .tiff files with the channels stored either before or after the image dimensions. A single-channel TIFF can also be a two-dimensional image.

For example, a dataset with wool and cotton classes can be organized as follows:

DatasetName/
├── wool/
│   ├── sample1.tif
│   └── sample2.tiff
└── cotton/
    ├── sample3.tif
    └── sample4.tif

If no train/validation split is provided, the platform automatically divides the samples into 70% training and 30% validation samples.

Providing a Pre-Split Dataset

If you prefer to define the training and validation sets yourself, place the same class folders inside separate train and val directories:

DatasetName/
├── train/
│   ├── wool/
│   │   └── sample1.tif
│   └── cotton/
│       └── sample2.tif
└── val/
    ├── wool/
    │   └── sample3.tif
    └── cotton/
        └── sample4.tif

Each class folder must contain at least one TIFF file. The class names in train and val must match, and the number of class folders must match the value entered during upload.

Dataset Registration

Once the folder structure is ready, compress the dataset as a .zip file. In the platform, open Collections, select Classification → Multispectral Imagery, and click Add a new multispectral classification dataset. Upload the archive, enter the number of classes, and set Number of channels to the number of bands in each TIFF file. The channel field accepts values from 1 to 64.

The platform checks the TIFF files and their channel count when the dataset is registered. After upload, the dataset appears in Collections, where you can view the samples, see a grayscale preview of the first band, and download the original TIFF files.

Training Initiation

After the dataset is uploaded, you can train a multispectral classification model:

  1. Navigate to the Train an ML model page.
  2. Select Classification → Multispectral Imagery.
  3. Choose a model and the uploaded collection.
  4. Select Fine-tuning (Pre-trained Model) or Train from Scratch, adjust the training parameters if needed, and click Start Training.

The platform uses the channel count registered with the dataset to configure the model. It calculates per-channel normalization statistics from the training images. The available augmentations are Blur, Scale, Rotate, and Flip. The best model is saved in Models according to its validation accuracy.

Inference Process

Once the model is trained, open it from Models to reach its inference page. Upload one or more .tif or .tiff files and click Start Inference. The TIFF files must have the same number and order of bands as the images used to train the model.

The results show a grayscale preview of the first band and the probability for each class. You can also download the original TIFF file.

Evaluation Metric

Accuracy measures the proportion of validation images assigned to the correct class. The platform uses this score to select the best model during training.