Real-Time Drought Monitoring And Prediction Using Multispectral Remote Sensing And Machine Learning

Abstract

This research examines real-time drought monitoring & prediction utilizing multispectral satellite data as well as machine learning algorithms.  Multiple drought indicators, including NDVI, VCI, TCI, LST, Rainfall, & SPI-3, were used to evaluate vegetative health and identify drought conditions in the Beed and Maharashtra areas.  The findings suggest that vegetative health was regularly poor between the years 2020 and 2023, with NDVI decreasing to 0.12, VCI and TCI falling below 40, through VHI reaching severe drought levels of 10-20, particularly between January and May.  Conditions only improved during the monsoon season, when rainfall climbed to 20-30 mm, bringing NDVI to 0.25-0.35 and VHI to 60-70. A drought-prediction model was created utilizing 262 Beed data and 2,355 Maharashtra samples, using VHI as the goal variable.  Beed's model has an average cross-validation score of R² = 0.977 ± 0.028 and outstanding overall accuracy (R² = 0.998, RMSE = 0.836).  The Maharashtra model performed much better, with CV R² = 0.993 ± 0.010, final R² = 0.999, & RMSE = 0.351.  Feature significance analysis revealed that vegetation-based variables dominated prediction, with VCI (0.43-0.46) and NDVI (0.41-0.47) accounting for more than 90% of the model's predictive power.  The temperature indicators (TCI = 0.03-0.05; LST = 0.03-0.06) had a little impact, whereas rainfall (~0.001) and SPI-3 (~0.000) provided nearly nothing. Overall, the data show that drought conditions in both locations are primarily caused by vegetative stress resulting from low rainfall along with rising land-surface temperatures.  The combination of multispectral remote sensing with machine-learning models allows a highly accurate and dependable method for real-time drought assessment, delivering a strong tool for agricultural planning and climate risk management in drought-prone regions.