Next-level vegetation health index forecasting: A ConvLSTM study using MODIS Time Series

Area of Science:

• Environmental Science
• Remote Sensing
• Machine Learning

Background:

• The Vegetation Health Index (VHI) is crucial for monitoring vegetation health using satellite data.
• Forecasting VHI is vital for agriculture and ecology, but advanced machine learning applications are limited.
• Existing methods often lack the precision needed for future vegetation health projections.

Purpose of the Study:

• To forecast Vegetation Health Index (VHI) values using remotely sensed images.
• To evaluate the effectiveness of a combined Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) model (ConvLSTM) for VHI forecasting.
• To compare traditional VHI calculation with a proposed global scale method using NDVI and LST.

Main Methods:

• Calculated VHI time series using Normalized Difference Vegetation Index (NDVI) and Land Surface Temperature (LST) data from MODIS.
• Employed a Convolutional Neural Network (CNN) and Long Short-Term Memory (LSTM) hybrid model (ConvLSTM) for time series forecasting.
• Utilized a novel global scale approach for VHI calculation, incorporating global minimum and maximum NDVI and LST values.

Main Results:

• A 1-layer ConvLSTM structure generally outperformed 2-layer and 3-layer models for VHI forecasting.
• Achieved low average Root Mean Square Error (RMSE) values: 0.025 (1-step), 0.026 (2-step), and 0.026 (3-step ahead).
• The global scale VHI calculation method, combined with ConvLSTM, demonstrated superior performance over traditional methods.

Conclusions:

• The ConvLSTM model, particularly a 1-layer structure, is effective for forecasting VHI from satellite imagery.
• The proposed global scale VHI calculation enhances forecasting accuracy.
• This approach offers a promising tool for advanced vegetation health monitoring and prediction.