Predictive Modeling of Vegetative Drought Using ML/DL Approach on Temporal Satellite Data

Abstract

The contemporary drought monitoring approaches are bounded by the need for greater visibility toward potentially hazardous scenarios. Hence, a temporal predictive analysis is aimed in this paper, which will be highly advantageous in subsequent planning for catastrophe mitigation and for presaging the vegetative health or probable drought event. Furthermore, the well-established Machine Learning (ML) models, comprising Random Forest and Ridge regressor, in addition to Deep Learning (DL) models, such as Multilayer Perceptron, 1D-CNN, and Pix2Pix Generative Adversarial Networks (P2P), are implemented across several timeframes of 1, 3, 6, 9, and 12 months. Also, the ML/DL models are trained by utilizing the Vegetative Health Index (VHI) values derived from NOAA/AVHRR satellite data from 1981 to 2022, with the Indian state of Karnataka conforming as the research region. In addition to generating temporal forecasts, the P2P model is further executed to perform an annual seasonal analysis that depicts the variations in dryness over time, Subsequently, the prediction performance is assessed through Mean Squared Error (MSE), Mean Absolute Error (MAE), and Coefficient of Determination (R2) scores. The pattern of prediction accuracy annotated demonstrates more accurate forecasts for durations of one month (short term) with the best R2 score, MSE and MAE notching up to 0.88, 0.009, and 0.055, respectively; consequently, as hypothesized, escorted by a decline with widening temporal gaps for future projections such as the yearly level (long term) where the R2 score, MSE, and MAE reduced up to 0.60, 0.030 and 0.114 respectively. Also, the seasonal analysis delivered valuable insights into the influences of various climatic factors on the dryness level of the landmass, which will act conducive to better future planning and preparation.