[Paper Review] A Machine Learning Approach to Long-Term Drought Prediction using Normalized Difference Indices Computed on a Spatiotemporal Dataset

Veronica Wairimu Muriga, Benjamin E. Rich+3 more|arXiv (Cornell University)|Feb 5, 2023

Hydrology and Drought AnalysisEnvironmental Science3 citations

TL;DR

This paper proposes a machine learning framework using Time-Distributed Convolutional Neural Networks (TD-CNNs) to predict the Normalized Difference Drought Index (NDDI) from spatiotemporal Sentinel-2 data in the SEN2DWATER dataset. It compares early-stage (direct NDDI training) and late-stage (NDVI and NDMI prediction followed by NDDI computation) NDDI prediction, finding that early computation yields significantly better performance with lower MSE (0.00054) and higher SSIM (0.99995) and PSNR (39.754).

ABSTRACT

Climate change and increases in drought conditions affect the lives of many and are closely tied to global agricultural output and livestock production. This research presents a novel approach utilizing machine learning frameworks for drought prediction around water basins. Our method focuses on the next-frame prediction of the Normalized Difference Drought Index (NDDI) by leveraging the recently developed SEN2DWATER database. We propose and compare two prediction methods for estimating NDDI values over a specific land area. Our work makes possible proactive measures that can ensure adequate water access for drought-affected communities and sustainable agriculture practices by implementing a proof-of-concept of short and long-term drought prediction of changes in water resources.

Motivation & Objective

• To develop a machine learning approach for long-term drought prediction using normalized difference indices derived from satellite imagery.
• To evaluate the effectiveness of Time-Distributed Convolutional Neural Networks (TD-CNNs) in predicting the Normalized Difference Drought Index (NDDI) across spatiotemporal sequences.
• To compare two NDDI computation strategies—early-stage (direct NDDI training) and late-stage (NDVI and NDMI prediction followed by NDDI calculation)—for predictive accuracy.
• To assess the feasibility of using the SEN2DWATER dataset as a benchmark for drought prediction using deep learning.
• To explore the potential of such models for proactive water resource management and climate resilience in agriculture and livestock systems.

Proposed method

• Utilizes the SEN2DWATER dataset, a spatiotemporal collection of Sentinel-2 multispectral images (13 bands) from lakes and rivers in Spain (2016–2022), with 39 cloud-minimized images per location over 6.5 years.
• Computes three indices: NDVI = (NIR - Red)/(NIR + Red), NDMI = (NIR - SWIR)/(NIR + SWIR), and NDDI = (NDVI - NDMI)/(NDVI + NDMI), using bands 8 (NIR), 4 (Red), and 11 (SWIR).
• Applies a Time-Distributed Convolutional Neural Network (TD-CNN) architecture with shared 2D convolutional layers across temporal frames and a subsequent ConvLSTM layer to model spatiotemporal dependencies.
• Trains two distinct prediction pipelines: (1) early-stage—training a single TD-CNN on pre-computed NDDI images; (2) late-stage—training separate TD-CNNs on NDVI and NDMI, then combining predictions to compute NDDI.
• Divides 300×300px images into 16×64×64px patches for training and validation, ensuring spatial generalization across multiple locations.
• Evaluates models using MSE, SSIM, and PSNR on test data, comparing performance across NDVI, NDMI, and both NDDI prediction methods.

Experimental results

Research questions

• RQ1 Can TD-CNNs effectively predict the NDDI index for long-term drought monitoring using spatiotemporal satellite data?
• RQ2 How does early-stage NDDI computation (training on pre-computed NDDI) compare to late-stage NDDI computation (combining NDVI and NDMI predictions) in terms of prediction accuracy?
• RQ3 What is the impact of data source characteristics—specifically, non-drought-prone regions like those in the SEN2DWATER dataset—on the performance of NDDI-based prediction models?
• RQ4 To what extent does the NDDI index suffer from resolution loss in deep learning-based prediction compared to NDVI and NDMI?
• RQ5 Can such models be used to reconstruct data under adverse conditions (e.g., cloud cover) or support multi-step forecasting in water resource management?

Key findings

• The early-stage NDDI prediction method achieved the lowest Mean Squared Error (MSE) of 0.00054, significantly outperforming the late-stage method (MSE = 0.0062).
• The early-stage NDDI model achieved a Structural Similarity Index (SSIM) of 0.99995 and a Peak Signal-to-Noise Ratio (PSNR) of 39.754, indicating near-perfect reconstruction fidelity.
• The late-stage NDDI prediction method, while still accurate, showed a noticeable drop in resolution and higher error, likely due to error propagation from two separate models.
• NDVI and NDMI predictions were highly accurate, with MSE values of 0.00743 and 0.01065, respectively, and high SSIM and PSNR scores, confirming their suitability for short-term prediction.
• The study suggests that the NDDI index may be less suitable for long-term prediction in non-drought-prone regions, as the SEN2DWATER dataset is derived from relatively stable, non-arid areas.
• The results support the feasibility of using SEN2DWATER as a benchmark for deep learning-based water resource prediction and highlight the need for improved resolution in NDDI modeling or alternative indices like SPEI for better performance.