IT-Drought: Interpretable Drought Forecasting

• The paper introduces IT-Drought, a Transformer framework that predicts daily root-zone soil moisture with subsequent flash drought diagnosis from physical state trajectories.
• It integrates an encoder-decoder architecture with lag-resolved Integrated Gradients to reveal dynamic hydroclimate memory and quantify key driver contributions.
• Empirical results show skillful forecasts up to 42 days and improved flash drought detection, outperforming traditional operational systems.

Interpretable Transformer for Drought (IT-Drought) denotes a Transformer-based drought forecasting framework in which the primary learned target is a physically meaningful hydrologic state—root-zone soil moisture—and flash drought is diagnosed from the predicted soil-moisture trajectory rather than being directly classified during training. In its named form, IT-Drought was introduced for daily forecasting across the contiguous United States (CONUS), combining an encoder-decoder Transformer with lag-resolved post hoc attribution via Integrated Gradients (IG) to support both operational prediction and scientific interpretation of flash drought onset (Zhao et al., 2 Oct 2025).

1. Definition and problem formulation

IT-Drought is posed as a subseasonal-to-seasonal hydroclimate forecasting system over CONUS. The direct machine-learning task is daily prediction of root-zone soil moisture over the $0\text{–}100$ cm layer. Flash drought forecasting is then performed as a downstream diagnostic step applied to the forecast soil-moisture sequence. This design is central to the method’s interpretability: the model predicts a physical state variable rather than a drought index or event label, and the drought diagnosis inherits the structure of the predicted state trajectory (Zhao et al., 2 Oct 2025).

The forecasting window uses a 365-day historical lookback and predicts the next 46 days simultaneously. Inputs and targets are represented at daily temporal resolution and $0.125^\circ$ spatial resolution. The study evaluates lead times from 1 to 46 days. The event definition for flash drought is procedural and based on weekly averaged root-zone soil-moisture percentiles computed separately for each grid cell and calendar week. A flash drought begins when weekly soil moisture falls below the 40th percentile, continues declining to the 20th percentile or lower, and must exhibit an average decline rate exceeding 5%. The onset phase ends when soil moisture starts to increase or the average decline rate drops below 5%, and the event terminates when soil moisture rises above the 20th percentile. Total event duration must be 1 to 3 months (Zhao et al., 2 Oct 2025).

This framing distinguishes IT-Drought from direct drought-classification systems. It also distinguishes it from longer-horizon binary PDSI classification benchmarks, such as monthly EarthFormer and ConvLSTM forecasting out to 12 months, and from weekly multi-index S2S drought forecasting frameworks that jointly predict surface soil moisture, ESI, and SIF. In IT-Drought, the physical target and the event-detection rule are tightly coupled, so forecast skill and interpretability are both anchored in soil-moisture dynamics rather than in a purely categorical label space (Marusov et al., 2023).

2. Architecture, inputs, and data representation

IT-Drought uses a standard Transformer with an encoder-decoder structure. Dynamic time series are passed through a linear transformation and positional encoding before entering the Transformer encoder. Static features are processed using multi-layer perceptrons and embedding techniques to learn spatial representations. The learned temporal and spatial representations are concatenated at each timestamp and then passed to the Transformer decoder, which generates the 46-day target root-zone soil-moisture sequence (Zhao et al., 2 Oct 2025).

The paper states that the model ingests 11 dynamic features and 7 static features, although the provided text explicitly names 9 dynamic predictor categories: precipitation, 2-m air temperature, surface pressure, vapor pressure deficit, wind speed, snowfall, shortwave incoming radiation, surface soil moisture $(0\text{–}40\ \mathrm{cm})$, and root-zone soil moisture $(0\text{–}100\ \mathrm{cm})$. The static features are land cover type, topography, canopy height, and the long-term mean and standard deviation of both surface and root-zone soil moisture over 1979–2022. All dynamic features are aggregated to daily, $0.125^\circ$ resolution, and all static features are aggregated to the same grid (Zhao et al., 2 Oct 2025).

There is no explicit spatial attention or graph-based interaction mechanism in the described architecture. The system appears to operate grid-cell by grid-cell, with static features providing learned local spatial context. The exact positional-encoding implementation is not reported, and the paper does not specify the number of Transformer layers, heads, hidden dimensions, or decoder cross-attention internals. The canonical attention mechanism consistent with the reported “standard Transformer” description is

$$\mathrm{Attention}(Q, K, V)=\mathrm{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V.$$

The paper does not present this equation explicitly in the text provided, but it is the standard mechanism implied by the stated encoder-decoder Transformer architecture (Zhao et al., 2 Oct 2025).

The model’s data pipeline is built entirely from public datasets. Root-zone and surface soil moisture come from NLDAS-2, precipitation from MSWEP v2.80, and the remaining climate variables from ERA5-Land. Static predictors come from the National Land Cover Database, Shuttle Radar Topography Mission, Global Forest Canopy Height 2019, and long-term NLDAS-2 climatologies. This public-data design is operationally relevant because it makes near-real-time deployment plausible without reliance on proprietary sensors or site-specific calibration (Zhao et al., 2 Oct 2025).

3. Training protocol, benchmarks, and predictive performance

The training protocol is temporally stratified. Training uses 32 years (1979–1994 and 2001–2016), validation uses 4 years (1995–1996 and 2017–2018), and testing uses 8 years (1997–2000 and 2019–2022). The study introduces a 2-year gap between training and testing periods to reduce contamination from soil-moisture memory. All inputs and the target are normalized to $[0,1]$ using min-max normalization. Reported optimization details include dropout $=0.1$, batch size $=32$, 2900 batches, and mean absolute error (MAE) loss on soil moisture. The paper does not report optimizer type, learning rate, number of epochs, or hardware (Zhao et al., 2 Oct 2025).

The benchmark compares IT-Drought against five operational process-based forecast systems using reforecasts: GEFS, ECMWF, CMA, CNRM, and HMCR. Soil-moisture skill is evaluated primarily with anomaly correlation coefficient (ACC), while flash-drought skill is evaluated with F1 score; precision and recall are also reported in extended results. The ACC at lead time $l$ is defined as

$0.125^\circ$0

and flash-drought performance uses the standard F1 score (Zhao et al., 2 Oct 2025).

A concise summary of the reported experimental configuration is given below.

Aspect	Specification	Notes
Input history	365 days	Daily dynamic inputs
Forecast horizon	46 days	Daily root-zone SM sequence
Spatial resolution	$0.125^\circ$1	CONUS grid
Training target	Root-zone SM	Flash drought derived afterward
Normalization	Min-max to $0.125^\circ$2	Applied to inputs and target
Loss	MAE of SM	Regression training
Split	Temporal	Includes 2-year gap

Empirically, IT-Drought achieves average CONUS ACC declining from $0.125^\circ$3 at 1 day lead to $0.125^\circ$4 at 46 days lead. The paper uses ACC $0.125^\circ$5 as the “skillful” threshold and reports that IT-Drought remains skillful to 42 days on average, whereas the best operational baseline remains skillful only to about 10 days. Existing forecast systems span roughly $0.125^\circ$6 to $0.125^\circ$7 ACC over 1–46 day leads. For flash drought, existing systems remain below $0.125^\circ$8 F1 at all lead times, while IT-Drought maintains F1 $0.125^\circ$9 up to 11 days lead on average across CONUS. Precision ranges from $(0\text{–}40\ \mathrm{cm})$0 to $(0\text{–}40\ \mathrm{cm})$1 and recall from $(0\text{–}40\ \mathrm{cm})$2 to $(0\text{–}40\ \mathrm{cm})$3 over 1–10 day leads. Over 1–14 day average lead times, $(0\text{–}40\ \mathrm{cm})$4 of CONUS has skillful soil-moisture forecasts with average ACC $(0\text{–}40\ \mathrm{cm})$5, and $(0\text{–}40\ \mathrm{cm})$6 has skillful flash-drought forecasts with average F1 $(0\text{–}40\ \mathrm{cm})$7 (Zhao et al., 2 Oct 2025).

The paper also reports that flash-drought forecasting becomes much weaker beyond about 3 weeks. F1 drops below $(0\text{–}40\ \mathrm{cm})$8, approaching process-based systems that lie around $(0\text{–}40\ \mathrm{cm})$9. This lead-time dependence is interpreted as a real predictability limit associated with rapid, high-frequency hydroclimate variability rather than as a purely architectural shortcoming (Zhao et al., 2 Oct 2025).

4. Interpretability framework and scientific findings

The interpretability layer in IT-Drought is based on Integrated Gradients rather than attention visualization. This is methodologically important: the model is Transformer-based, but the explanation method is not the attention matrix. The IG formulation given in the paper is

$(0\text{–}100\ \mathrm{cm})$0

where $(0\text{–}100\ \mathrm{cm})$1 is the trained soil-moisture forecasting network, $(0\text{–}100\ \mathrm{cm})$2 is the observed input, and the baseline for dynamic features is a zero vector (Zhao et al., 2 Oct 2025).

IG is applied during flash-drought onset. For each onset day, the trained model receives the prior 365-day history plus static features; IG is then computed for each feature at each lag day, aggregated across onset cases, and converted into lag-resolved attribution curves. The resulting analysis identifies both contribution magnitude and effective time range for antecedent drivers. Root-zone soil moisture self-dependency is the dominant positive contributor to forecasted root-zone soil moisture during flash-drought onset, and its contribution is reported to be about one order of magnitude larger than those of climate variables. Surface soil moisture also contributes materially through infiltration (Zhao et al., 2 Oct 2025).

The paper defines the effective soil-moisture-memory range $(0\text{–}100\ \mathrm{cm})$3 as the number of antecedent days for which root-zone soil-moisture IG exceeds $(0\text{–}100\ \mathrm{cm})$4. Under this criterion, mean $(0\text{–}100\ \mathrm{cm})$5 across CONUS is $(0\text{–}100\ \mathrm{cm})$6 days. Memory is longer in the Northeast, Midwest, and Northwest and shorter in the Southeast and Southwest. The study further reports that $(0\text{–}100\ \mathrm{cm})$7 positively correlates with long-term average root-zone soil moisture, and interprets this as evidence that wetter regions, with larger soil-water storage and longer residence times, retain longer predictive memory (Zhao et al., 2 Oct 2025).

Among climate variables, high antecedent shortwave radiation is the strongest drying driver, with the largest climate IG magnitude and strongest effect over roughly the previous week. High air temperature also reduces soil moisture, with a more persistent effect and an average effective time range of 9 days, particularly in summer and autumn. Low antecedent precipitation contributes to flash drought as well, especially in winter and spring and especially in topographically complex or snow-affected regions. High surface pressure has a drying effect with a short-lived effective range of around 3 days. High VPD has only minimal impact in humid regions but larger impact in southwestern drylands. Low wind speed and low snowfall also contribute regionally (Zhao et al., 2 Oct 2025).

These results supply a mechanistic vocabulary for an otherwise black-box sequence model. They imply that IT-Drought’s skill is rooted first in antecedent soil-moisture memory and only secondarily in climatic forcing, with radiation and temperature operating as the principal short-lag atmospheric drivers. The paper explicitly treats these as predictive contributions rather than causal estimates, so the interpretability claim is post hoc and attributional rather than experimental or intervention-based (Zhao et al., 2 Oct 2025).

5. Position within the drought-transformer literature

IT-Drought sits within a broader line of arXiv work that has progressively combined drought prediction with explicit explanatory machinery. Earlier monthly geospatial forecasting work benchmarked EarthFormer, ConvLSTM, FourCastNet, logistic regression, and XGBoost for binary PDSI forecasting at horizons $(0\text{–}100\ \mathrm{cm})$8 months across Missouri, Madhya Pradesh, Goiás, Northern Kazakhstan, and Poland. That study found EarthFormer strongest for short horizons, especially 1 and 3 months, while ConvLSTM dominated at 9 and 12 months, and it did not implement explicit interpretability methods. Its main relevance to IT-Drought is architectural: it establishes cuboid-based Transformer forecasting as effective for short- to medium-range drought prediction, while also showing that recurrent inductive bias remains advantageous for longer memory horizons (Marusov et al., 2023).

A separate benchmark, DroughtSet, framed S2S drought as weekly prediction of three continuous targets—surface soil moisture, ESI, and SIF—over CONUS at 4 km resolution using 100 weeks of context to forecast 26 future weeks. Its model, SPDrought, coupled a local static-conditioned spatial fusion module with a Transformer encoder-decoder and used Integrated Gradients plus leave-one-feature-out analysis for interpretation. SPDrought achieved total MAE $(0\text{–}100\ \mathrm{cm})$9, outperforming Transformer, Informer, PatchTST, DLinear, iTransformer, TimesNet, and LSTM, and achieved soil-moisture-drought detection accuracy $0.125^\circ$0 and precision $0.125^\circ$1 under a 30th-percentile threshold. Relative to IT-Drought, SPDrought is broader in drought targets and spatial fusion, but it is weekly rather than daily and does not use the soil-moisture-first flash-drought diagnosis strategy (Tan et al., 2024).

In crop-scale remote sensing, an explainable ViT-B/16 pipeline for potato drought stress detection used low-altitude aerial RGB imagery, transfer learning, and either a binary ViT classifier or a ViT feature extractor followed by an RBF-kernel SVM. The end-to-end ViT achieved test accuracy $0.125^\circ$2, while the ViT+SVM with optimal weights achieved 5-fold mean accuracy $0.125^\circ$3 and mean AUC $0.125^\circ$4. Its interpretability mechanism was attention-map visualization across all 12 encoder blocks. That study shows a different, image-centric usage of interpretable drought Transformers: localizing spatial drought signatures in crop canopies rather than forecasting hydroclimate states over time (Patra et al., 2024).

Hydrometeorological explainability work outside drought has also informed this trajectory. A Transformer variant for daily water-level forecasting integrated sparse attention and a nonlinear decoder output layer and used SHAP for both global and local feature attribution, identifying dominant contributions from hydrological state and temperature variables. The model’s best variant, Transformer-EN, achieved $0.125^\circ$5 at 1-day lead and remained strongest through 7-day forecasting. Although the target is water level rather than drought, the paper demonstrates a viable pattern: Transformer prediction plus post hoc XAI rather than attention-as-explanation (Liu et al., 2024).

Finally, physics-guided drought-impact modeling has introduced a complementary notion of interpretability. One study modeled crop yield as a function of temporal water scarcity through learned $0.125^\circ$6 and $0.125^\circ$7, enforced by a physics-informed loss and estimated at $0.125^\circ$8 m resolution with deep ensembles; another embedded a LINTUL5 biomass-growth law into a neural network and recovered latent physiological variables $0.125^\circ$9, defining drought days as $$\mathrm{Attention}(Q, K, V)=\mathrm{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V.$$0. These systems are not named IT-Drought and are not Transformer-based in their primary form, but they show that drought interpretability can also be structural and process-informed rather than purely attributional (Miranda et al., 21 Oct 2025, Shi et al., 22 Jan 2026).

6. Limitations, misconceptions, and open directions

Several limitations delimit the current meaning of IT-Drought. First, the model’s interpretability is post hoc rather than intrinsic. It is “interpretable” because it predicts a physically meaningful target and is analyzed with IG, not because the Transformer architecture itself is self-explanatory. A common misconception is therefore to equate IT-Drought with attention visualization. The paper does not use attention as its explanation method; it explicitly relies on IG (Zhao et al., 2 Oct 2025).

A second misconception is that IT-Drought is a flash-drought classifier. It is not. Training is single-task regression on root-zone soil moisture, and flash drought is diagnosed afterward from the predicted sequence. This distinction matters methodologically because the model never sees drought labels as direct optimization targets, and no explicit class-imbalance correction is reported (Zhao et al., 2 Oct 2025).

A third limitation concerns scope and reproducibility. The study is confined to CONUS and does not test transfer to other continents or climates. Forecast quality is benchmark-dependent because the primary reference is NLDAS-2, and process-based systems can appear stronger when they share components or initialization sources with the benchmark product. The paper also omits many architectural and optimization details, including layer counts, head counts, optimizer, and learning rate, which constrains exact replication from the manuscript alone. Long-lead flash-drought skill also degrades substantially beyond about 3 weeks, and the paper notes that some rapid soil-moisture variability arises from diurnal plant water use, atmospheric feedbacks, irrigation, and other human activities that are not fully represented in the inputs (Zhao et al., 2 Oct 2025).

Related work sharpens the open design space. EarthFormer-style local spatiotemporal attention appears effective for short- to medium-range monthly drought forecasting, but ConvLSTM remains stronger at 9–12 months, suggesting that stronger temporal inductive bias is still useful for long-horizon drought memory (Marusov et al., 2023). DroughtSet suggests that explicit spatial fusion, static context, and multi-target learning improve weekly S2S drought forecasting (Tan et al., 2024). Physics-guided crop-stress models suggest that future IT-Drought variants may benefit from interpretable intermediate states such as $$\mathrm{Attention}(Q, K, V)=\mathrm{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V.$$1, $$\mathrm{Attention}(Q, K, V)=\mathrm{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V.$$2, $$\mathrm{Attention}(Q, K, V)=\mathrm{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V.$$3, $$\mathrm{Attention}(Q, K, V)=\mathrm{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V.$$4, $$\mathrm{Attention}(Q, K, V)=\mathrm{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V.$$5, or $$\mathrm{Attention}(Q, K, V)=\mathrm{softmax}\left(\frac{QK^\top}{\sqrt{d_k}}\right)V.$$6, especially when the goal extends from hazard detection to agricultural impact assessment. This suggests a plausible next generation of IT-Drought systems in which Transformer sequence modeling is combined with physically constrained latent variables, uncertainty estimation, and lead-specific explanatory analysis rather than relying on attribution alone (Miranda et al., 21 Oct 2025, Shi et al., 22 Jan 2026).

In that broader sense, IT-Drought is best understood not merely as a single model instance but as a concrete formulation of a larger methodological shift in drought informatics: from direct index prediction toward physically meaningful state forecasting, and from opaque deep learning toward explanation layers that expose lag structure, driver hierarchy, and regional heterogeneity.