Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Long-term drought prediction using deep neural networks based on geospatial weather data

Alexander Marusov, Vsevolod Grabar, Yury Maximov, Nazar Sotiriadi, Alexander Bulkin, Alexey Zaytsev

TL;DR

This study targets long-term drought prediction by reframing PDSI-based drought assessment as a spatio-temporal classification task. It benchmarks baselines, ConvLSTM, and transformer-based models (FourCastNet, EarthFormer) across five global regions using TerraClimate-derived PDSI data, demonstrating that EarthFormer excels at short horizons while ConvLSTM provides superior long-horizon forecasts. The results show significant improvements over traditional methods and highlight the value of ensembling and region cropping to boost predictive reliability. Overall, the work advances practical drought forecasting for agriculture and insurance with an end-to-end, data-efficient deep-learning framework and broad regional validation.

Abstract

The problem of high-quality drought forecasting up to a year in advance is critical for agriculture planning and insurance. Yet, it is still unsolved with reasonable accuracy due to data complexity and aridity stochasticity. We tackle drought data by introducing an end-to-end approach that adopts a spatio-temporal neural network model with accessible open monthly climate data as the input. Our systematic research employs diverse proposed models and five distinct environmental regions as a testbed to evaluate the efficacy of the Palmer Drought Severity Index (PDSI) prediction. Key aggregated findings are the exceptional performance of a Transformer model, EarthFormer, in making accurate short-term (up to six months) forecasts. At the same time, the Convolutional LSTM excels in longer-term forecasting.

Long-term drought prediction using deep neural networks based on geospatial weather data

TL;DR

This study targets long-term drought prediction by reframing PDSI-based drought assessment as a spatio-temporal classification task. It benchmarks baselines, ConvLSTM, and transformer-based models (FourCastNet, EarthFormer) across five global regions using TerraClimate-derived PDSI data, demonstrating that EarthFormer excels at short horizons while ConvLSTM provides superior long-horizon forecasts. The results show significant improvements over traditional methods and highlight the value of ensembling and region cropping to boost predictive reliability. Overall, the work advances practical drought forecasting for agriculture and insurance with an end-to-end, data-efficient deep-learning framework and broad regional validation.

Abstract

The problem of high-quality drought forecasting up to a year in advance is critical for agriculture planning and insurance. Yet, it is still unsolved with reasonable accuracy due to data complexity and aridity stochasticity. We tackle drought data by introducing an end-to-end approach that adopts a spatio-temporal neural network model with accessible open monthly climate data as the input. Our systematic research employs diverse proposed models and five distinct environmental regions as a testbed to evaluate the efficacy of the Palmer Drought Severity Index (PDSI) prediction. Key aggregated findings are the exceptional performance of a Transformer model, EarthFormer, in making accurate short-term (up to six months) forecasts. At the same time, the Convolutional LSTM excels in longer-term forecasting.
Paper Structure (34 sections, 7 figures, 12 tables)

This paper contains 34 sections, 7 figures, 12 tables.

Table of Contents

  1. Highlights
  2. Software and data availability
  3. CRediT author statement
  4. Introduction
  5. Data
  6. Methods
  7. Baseline
  8. Basic methods: Logistic regression and Gradient boosting
  9. Logistic Regression
  10. Gradient boosting
  11. Convolutional LSTM (ConvLSTM)
  12. Details of architecture
  13. Transformer-based methods
  14. FourCastNet
  15. EarthFormer
  16. ...and 19 more sections

Figures (7)

  • Figure 1: Diverse regions chosen for PDSI forecast
  • Figure 2: Our version of ConvLSTM architecture
  • Figure 3: Quality metrics for the binary drought severity classification: median $\mathrm{ROC\, AUC}$, $\mathrm{PR\, AUC}$, and $\mathrm{F1}$ for different forecast horizons averaged over five considered regions
  • Figure 4: Spatial distribution of $\mathrm{ROC\, AUC}$ for 6 month forecast, Missouri, ConvLSTM
  • Figure 5: Evolution of spatial maps with zooming in: top to bottom area decreases from 100% of the region to 27%, left to right forecast horizons are 1, 3, 6, and 12 months correspondingly.
  • ...and 2 more figures