Overview

PhenoNN is a deep learning framework for predicting phenological metrics from climate data. It uses LSTM (Long Short-Term Memory) networks to capture temporal dependencies in climate time series and predict Green Chromatic Coordinate (GCC) for different plant functional types.

Architecture

The PhenoNN architecture consists of:

  1. Data Preprocessing: Climate data is standardized/normalized and formatted for LSTM input

  2. LSTM Network: Multi-layer LSTM with configurable hidden size and dropout

  3. Training Pipeline: Cross-validation, early stopping, and learning rate scheduling

  4. Prediction Pipeline: Ensemble predictions from multiple trained models

Plant Functional Types

PhenoNN supports three plant functional types:

  • DB (Deciduous Broadleaf): Deciduous broadleaf forests

  • EN (Evergreen Needleleaf): Evergreen needleleaf forests (coniferous)

  • GR (Grassland): Grasslands and herbaceous vegetation

Feature Sets

The model can use different combinations of features:

  • 6 features: Dynamic variables only (temperature, precipitation, radiation, etc.)

  • 8 features: Dynamic + static climate (mean annual temperature and precipitation)

  • 9 features: Dynamic + static + snow cover

  • 14 features: All available variables (dynamic, static, soil properties)

Input Data Format

Input data should be in CSV format with daily climate variables for two years to predict one year of GCC observations. Required variables include:

Dynamic Variables: - Daily minimum temperature (tmin) - Daily maximum temperature (tmax) - Daily daylength (daylength) - Daily vapor pressure deficit (vpd) - Daily soil water availability (swa) - Daily