Models

Single-task CNN model

A multiple-regression convolutional neural network for each predicted trait. Each model is comprised of many one-dimensional convolution layers, followed by fully-connected layers.

Multi-task CNN model

A multivariate regression convolutional neural network that predicts all trait values concurrently. Like the single-task CNN, it is comprised of many one-dimensional convolution layers, followed by fully-connected layers to make a prediction for each trait.

PLS model

An individual partial least squares regression model per trait, which operates on the reflectance measurements alone. Due to the nature of PLS, this only works on fixed-length inputs (and as such, a few models will be made available for a selection of spectral ranges).

As the PLS model doesn’t have the capacity to make predictions for variable spectral ranges, a small number of separate models have been built using commonly-encountered wavelength ranges (listed below). The uploaded reflectance data will be trimmed to match PLS model with the most-similar spectral range. If the CSV data cannot be trimmed to match a model, it will be rejected by the server.

PLS Spectral ranges:

• [400, 900]
• [400, 1000]
• [400, 1700]
• [400, 2400]

Ensemble

An ensemble composed of the PLS, CNN-multi, and CNN-single models. Each of these models makes a trait prediction, and the ensemble prediction is computed as the mean of these.