Modeling time series of vegetation indices in tallgrass prairie using machine and deep learning algorithms

Abstract

The vegetation phenology of tallgrass prairie varies yearly, depending on climatic conditions, plant species composition, and location. Modeling time series of vegetation indices (VIs) using climate data can be useful for understanding and predicting how tallgrass prairie will respond to future climate scenarios and for identifying and managing areas of tallgrass prairie that are particularly susceptible to climate-induced changes. Machine or deep learning algorithms can be well-suited to model VIs for phenology studies by identifying patterns and relationships between climatic factors and VIs using historical data. This study evaluated the performance of 12 machine and deep learning algorithms, encompassing a diverse range of algorithmic families, in modeling patterns of the Moderate Resolution Imaging Spectroradiometer-derived enhanced vegetation index (EVI, greenness index) and land surface water index (LSWI) in native tallgrass prairie. The models include linear regression, Bayesian ridge, elastic net, decision tree, random forest, eXtreme Gradient Boosting (XGBoost), support vector regression (SVR), K-nearest neighbors (KNN), artificial neural network (ANN), convolutional neural network (CNN), recurrent neural network (RNN), and long short-term memory (LSTM). Air and soil temperatures showed the highest correlations with EVI (r ≥ 0.77) and LSWI (r ≥ 0.56). The low correlation (r ≤ 0.23) of EVI and LSWI with contemporaneous rainfall or soil moisture suggests vegetation's delayed response to these factors. The results indicated that ensemble methods like XGBoost and random forest performed best across all three datasets (i.e., training, testing, and validation) for modeling EVI and LSWI. Deep learning models showed varying performance across datasets, and their performance was sub-optimal compared to XGBoost and random forest. The linear regression also showed a moderate performance, while the decision tree performed the weakest overall. The strong performance of XGBoost and random forest highlights the intricate and nonlinear relationship of prairie vegetation with climatic factors. These models' strength lies in capturing such complexities. This study provides insights into the key climatic factors and underlying processes that control the vegetation dynamics of tallgrass prairie ecosystems. Our machine learning models can be a valuable tool for developing new strategies to manage tallgrass prairie ecosystems in the face of climate change.