Forecasting Vegetation Behavior Based on PlanetScope Time Series Data Using RNN-Based Models

Abstract

Accurate vegetation behavior forecasting is essential for understanding the dynamics of plant life in the context of climate change and other natural or human-induced disturbances. Recurrent neural network (RNN) deep learning (DL) models represent a modern approach to predict vegetation behavior with a high level of precision. In this article, we explore the potential of different DL and more traditional methods to forecast the normalized difference vegetation index (NDVI), which is directly related to the state of vegetation and its dynamics. A time-series dataset consisting of 70 NDVI images calculated from PlanetScope data from April 2017 to January 2023 was used. Initially, all selected methods were evaluated and compared. From the six tested methods, simple RNN (SRNN) proved to be the most accurate method for predicting vegetation dynamics. The SRNN model results achieved a mean RMSE of 0.051 when compared to the actual 2022 NDVI values. The high accuracy was reflected in all five studied vegetation classes characterizing the selected Mediterranean test area. The SRNN method performs very well in most months, except in autumn where it underestimates NDVI values. To get a thorough insight into the results, we also compared them to the Sentinel-2 NDVI data and climate data consisting of temperature and precipitation values. It was found that most of the prediction differences were due to the irregular variations in meteorological conditions during the year analyzed. The predictive capabilities of RNNs are an effective tool for forecasting vegetation dynamics but can be further improved by incorporating climate data into the prediction process.

Keywords

• Climatic data
• deep learning (DL)
• normalized difference vegetation index (NDVI)
• satellite imagery
• spatio-temporal prediction
• vegetation dynamics