AI-Based Modeling of Post-Fire Evapotranspiration Using Vegetation Recovery Indicators: Application to the 2022 Chongqing Burned Areas

Abstract

The 2022 Chongqing wildfires, occurring during an unprecedented heatwave, severely degraded subtropical forest ecosystems and disrupted hydrological cycling. We developed an integrated artificial intelligence framework combining Long Short-Term Memory and Transformer architectures to simulate post-fire evapotranspiration (ET) dynamics using 37 months of field observations (2022–2025) across 24 plots with four burn severities. The Penman–Monteith–Leuning model provided physically based benchmarks. Results revealed three distinct recovery phases: destruction/stagnation (0–7 months, ET at 6%–10% of pre-fire levels), rapid recovery (8–19 months), and stabilization (20–37 months, reaching 100% ET recovery). The coupled LSTM–Transformer ensemble achieved superior performance (RMSE = 0.10 mm·day⁻¹, NSE = 0.98), outperforming single models by 31% in uncertainty reduction. SHAP analysis identified phase-dependent factor shifts: soil water content dominated Stage I (42.5%), while leaf area index (LAI) controlled Stages II–III (>48%). A bimodal LAI time-lag effect emerged: 4–7 days (leaf water potential equilibrium, 27.7% contribution) and 8–14 days (root uptake compensation, 21.7%). Burn severity significantly extended time-lags (severe burns: 12/21 days vs. unburned: 5/12 days), indicating hydraulic system reconstruction requirements. Despite equivalent LAI recovery, severe burns maintained 12%–15% ET reduction, suggesting lasting hydraulic limitations. This study demonstrates that physics-constrained AI models effectively capture complex post-fire ecohydrological dynamics while providing mechanistic interpretability, advancing understanding of vegetation–water coupling reconstruction under increasing fire frequency.

1. Introduction

Wildfires have intensified globally under climate change, with increasing frequency, severity, and extent across diverse biomes [1,2]. The 2022 summer witnessed unprecedented fire activity in Chongqing, Southwest China—a region historically considered too humid for large-scale wildfires. Extreme drought conditions, with 60% precipitation reduction and 40+ consecutive days exceeding 40 °C, created fire risk conditions outside the historical envelope, resulting in approximately 300 hectares of forest destruction across Beibei and Banan Districts [3,4]. This event exemplifies the “fire weather” phenomenon where climate change expands fire-prone regions into areas previously considered fire-safe [5,6].

Post-fire ecosystem recovery involves complex interactions between vegetation regeneration, soil rehabilitation, and hydrological cycling [7]. Evapotranspiration (ET), as the primary pathway of terrestrial water flux, serves as a critical indicator of ecosystem functional recovery, integrating canopy structure, root system integrity, and soil-vegetation-atmosphere coupling [8,9]. However, ET recovery trajectories remain poorly characterized in subtropical forest ecosystems, particularly following extreme fire events that cause complete canopy destruction.

Traditional ET estimation relies on physical models (Penman–Monteith, Priestley–Taylor) or remote sensing products (MOD16, PT-JPL), but these approaches face significant limitations in post-fire environments. Physical models require parameterized canopy conductance that changes dramatically during recovery [10], while satellite-derived ET suffers from cloud contamination, terrain shadows, and saturation effects in dense vegetation [11,12]. More critically, neither approach captures the nonlinear, time-lagged relationships between vegetation recovery and ET reestablishment.

Deep learning offers promising alternatives for complex ecohydrological modeling [2,13], with Long Short-Term Memory (LSTM) networks excelling at temporal dependency capture and Transformer architectures demonstrating superior performance in long-range sequence modeling. Recent studies have demonstrated the efficacy of deep learning for evapotranspiration (ET) modeling, with LSTM networks excelling in capturing long-term temporal dependencies in hydrological time series, while Transformer architectures leverage self-attention mechanisms to identify complex feature interactions across varying time scales. However, existing applications predominantly focus on agricultural systems or intact forest ecosystems under stable conditions [14]. However, AI applications in post-fire ET prediction remain scarce, and the “black box” nature of neural networks limits mechanistic interpretation essential for ecological understanding [15].

This study addresses these gaps by developing a physics-constrained AI framework for post-fire ET modeling with three specific objectives: (1) to evaluate LSTM, Transformer, and ensemble architectures against physical benchmarks across distinct recovery phases; (2) to quantify phase-dependent controlling factors and time-lag effects using SHAP interpretability analysis; and (3) to characterize burn severity impacts on ET-LAI coupling mechanisms. The 2022 Chongqing wildfires provide a unique case study of subtropical forest recovery from extreme fire, with implications for fire ecology and ecosystem management under climate change.

2. Materials and Methods

2.1. Information on the Study Area

This study was conducted in two distinct wildfire-affected regions (Figure 1) within Chongqing Municipality, Southwest China, both of which experienced severe forest fires during the unprecedented extreme heatwave conditions of August 2022. The first study site was located in Beibei District, specifically at Xiema Street (29°50′ N, 106°24′ E). The experimental site is located on the slopes of Jinyun Mountain within the district. The fire at this location ignited at approximately 22:30 on 21 August 2022, and the vegetation community was characterized as subtropical evergreen broad-leaved forest dominated by Castanopsis fargesii and Lithocarpus glaber with scattered occurrences of Pinus massoniana, representing the typical zonal vegetation of this region. The second study site was situated in Banan District, specifically within Jieshi Town (29°15′ N, 106°35′ E), where fire ignition occurred earlier on the same day at approximately 17:00. The vegetation at the Banan site was classified as mixed coniferous-broadleaf forest with co-dominance of Pinus massoniana and Quercus variabilis, representing a slightly more drought-prone forest type compared to the Beibei site. Both locations experienced extreme drought conditions throughout 2022, with precipitation reduced by approximately 60% compared to long-term averages and more than 40 consecutive days recording temperatures exceeding 40 degrees Celsius, creating fire risk conditions unprecedented in the historical climate record of this typically humid subtropical region. The wildfires at both sites were suppressed through intensive firefighting efforts within 5 to 7 days of ignition but caused complete canopy destruction across significant portions of the burned areas, with the Beibei fire consuming approximately 180 hectares and the Banan fire affecting approximately 120 hectares of forest vegetation [3,16,17]. Pre-fire forests were 35–45-year-old secondary evergreen broadleaf stands on acidic yellow mountain soils (pH 5.2–6.5, 30–45 cm depth) derived from purple sandstone, with 1.2–1.8 g·kg⁻¹ total nitrogen, established following 1980s agricultural abandonment and unmanaged for two decades prior to the 2022 extreme drought-fire event.

2.2. Experimental Design and Field Measurements

At each site, 12 permanent 30 m × 30 m plots were established along a continuous burn-severity gradient, classified using the differenced Normalized Burn Ratio (NBR) from Sentinel-2. Four severity classes were defined: unburned (NBR < 0.1, intact canopy), low (0.10–0.27, 30%–50% scorch), moderate (0.27–0.44, 50%–80% canopy consumption), and high (>0.44, >80% consumption, near-total mortality). Three replicate plots per severity class were installed at each site (a total of 24 plots), georeferenced with differential GPS, and permanently marked. Leaf Area Index (LAI), the primary recovery indicator, was measured monthly from the pre-fire baseline (August 2022) to August 2025 (37 months) using an LAI-2200C plant canopy analyzer (LI-COR, Inc., Lincoln, NE, USA). The 24 plots were located within the 80% cumulative EC footprint zone, with LAI measured at plot centers and four cardinal directions, then upscaled to footprint scale via wind-weighted averaging based on dominant northeast and southwest wind patterns. Measurements were taken under diffuse light (overcast or dawn/dusk); five below-canopy readings were averaged per plot. Cross-calibration with digital hemispherical photography showed strong agreement (R² = 0.94). Replicate averages minimized microsite variability. Eddy covariance systems (CSAT3A sonic anemometer and EC150 gas analyzer, Campbell Scientific, Inc., Logan, UT, USA) were deployed at unburned, moderate-, and high-severity plots per site, mounted at 3.5 m. Raw 10 Hz data were processed with EddyPro (version 7.0.6, LI-COR Biosciences, Lincoln, NE, USA) (friction-velocity threshold > 0.15 m·s⁻¹); energy-balance closure was corrected via Bowen ratio. The 3-PG2 model showed strong pre-fire performance but required parameter adjustments for post-fire conditions, with limited reliability during initial severe canopy disruption when direct measurements were prioritized over gap-filling.

2.3. ET Calculation and Modeling Framework

The Penman–Monteith–Leuning (PML) model provided physics-based ET benchmarks for AI model training, computing latent heat flux from available energy, vapor pressure deficit (VPD), aerodynamic conductance, and stomatal conductance modulated by radiation, VPD, soil moisture, and temperature [18]. Vegetation cover fraction derived from LAI used an exponential extinction coefficient of 0.5. Post-fire parameters were adjusted: maximum stomatal conductance was reduced by 80%, and soil albedo increased by 15% for high-severity conditions, gradually restoring with LAI recovery. Deep learning models predicted ET from ten daily variables: air temperature, VPD, solar radiation, wind speed, precipitation, soil moisture (10 cm depth), measured and gap-filled LAI from the 3-PG2 model, LAI change rate, and day-of-year encoding. Meteorological data were collected using Campbell Scientific CR1000X weather stations (Campbell Scientific, Inc., Logan, UT, USA) installed within each plot, recording air temperature, precipitation, relative humidity, wind speed, and solar radiation at 10 min intervals. Output was daily ET (mm·day⁻¹), with simulation and 1–30 day prediction modes. The LSTM architecture used two bidirectional layers (64 and 32 hidden units) with attention and a 16-neuron dense layer, employing 30–90 day input windows. The Transformer employed the Informer architecture with self-attention, multi-scale temporal encoding, and LAI-guided cross-attention. The coupled ensemble combined LSTM and Transformer predictions with Bayesian-optimized equal weighting and physics-constrained loss functions penalizing water balance violations. Training used temporal cross-validation with expanding windows and cosine annealing learning rates (0.001 to 0.00001). Early stopping was triggered when validation loss plateaued for 10 epochs. Loss combined mean squared error with physical constraint terms on water balance and theoretical ET-LAI relationships. The Informer architecture was selected for its ProbSparse attention mechanism, which, superiorly, captures long-range dependencies in recovery dynamics, while Bayesian optimization converged to near-equal ensemble weights (LSTM: 0.48, Informer: 0.52), indicating comparable model contributions, where complex weighting schemes yielded no performance improvement.

2.4. Model Evaluation and Interpretability Analysis

Model performance was evaluated using multiple complementary metrics, including root mean square error for absolute error quantification, Nash-Sutcliffe efficiency for assessment of explained variance relative to the mean, Kling-Gupta efficiency for integrated evaluation of correlation, variability, and bias components, mean bias error for systematic deviation detection, and coefficient of determination for overall explanatory power. These metrics were computed for each model across all burn severity classes and recovery stages, with particular attention to performance differences between the destruction and stagnation phase, rapid recovery phase, and stabilization phase of ecosystem restoration. Input features (X) comprised meteorological variables and LAI, while the target variable (y) was actual evapotranspiration measured directly by Eddy Covariance (EC) systems. The Penman–Monteith–Leuning model generated reference estimates solely for comparative validation, not as training targets. EC measurements were validated through energy balance closure analysis (closure ratio: 0.82–0.91), with predictions assessed against these observed values using Nash-Sutcliffe efficiency and RMSE.

Model interpretability was achieved through SHAP analysis to quantify the contribution of each input feature to ET predictions, with global importance calculated as mean absolute SHAP values across all predictions, time-lag decomposition computed for LAI effects at daily intervals from zero to 30 days prior to identify delayed responses, interaction effects evaluated for feature pairs to reveal synergistic controls, and severity-stratified analysis conducted separately for unburned, low severity, moderate severity, and high severity classes to detect burn-dependent shifts in controlling mechanisms [19]. All statistical analyses were implemented in Python (version 3.9.20) using TensorFlow (version 2.16.1) for deep learning, the SHAP library for interpretability analysis, and SciPy (version 1.13.1) for significance testing, with two-way analysis of variance employed for model-by-stage comparisons and Tukey honestly significant difference tests for post hoc pairwise comparisons at the 0.05 significance level.

3. Results

3.1. Post-Fire LAI Dynamics and ET Recovery Patterns

Following the August 2022 wildfire in Chongqing, the severely burned areas exhibited dramatic vegetation degradation, with LAI declining from 7.7 m²·m⁻² (pre-fire) to near-zero values (0.1 m²·m⁻²) immediately post-fire (Figure 2). The recovery trajectory demonstrated distinct phase-dependent patterns over the 37-month observation period (August 2022–August 2025).

During Stage I (destruction and stagnation phase, August 2022–February 2023), LAI remained below 0.3 m²·m⁻² for seven months, indicating minimal vegetation reestablishment. Correspondingly, ET values were severely suppressed, ranging from 0.28 to 0.45 mm·day⁻¹ (

Table 1. Comparative ET simulation performance across post-fire recovery stages (August 2022–August 2025). ET values (mm·day⁻¹) from four models (PML, LSTM, Transformer, and Coupled) with corresponding LAI (m²·m⁻²), recovery stage classification, and restoration ratio (%). Jinyun Mountain severe burn plot is shown as a representative example.

Date	LAI m2·m−2	ET/PML mm·Day−1	ET/LSTM mm·Day−1	ET/Transformer mm·Day−1	ET/Coupled mm·Day−1	Stage	Restoration Ratio (%)
Aug-22	0.1	0.45	0.52	0.38	0.45	I	10
Oct-22	0.15	0.38	0.45	0.32	0.38	I	8.4
Dec-22	0.22	0.28	0.38	0.22	0.28	I	6.2
Feb-23	0.3	0.28	0.38	0.22	0.28	I→II	6.2
Apr-23	0.42	0.48	0.58	0.38	0.48	II	10.7
Jun-23	0.68	0.95	1.05	0.85	0.95	II	21.1
Jul-23	0.92	1.35	1.45	1.25	1.35	II	30
Aug-23	1.28	1.85	1.95	1.75	1.85	II	41.1
Oct-23	1.75	2.35	2.45	2.25	2.35	II	52.2
Dec-23	2.28	2.65	2.75	2.55	2.65	II→III	58.9
Mar-24	3.08	3.15	3.25	3.05	3.15	III	70
Jul-24	3.88	3.75	3.85	3.65	3.75	III	83.3
Dec-24	4.42	3.95	4.05	3.85	3.95	III	87.8
Jul-25	5.32	4.35	4.45	4.25	4.35	III	96.7
Aug-25	5.42	4.65	4.75	4.55	4.65	III	100

), representing only 6%–10% of pre-fire levels.

Stage II (rapid recovery phase, March 2023–February 2024) witnessed accelerated LAI regeneration, increasing from 0.35 to 2.28 m²·m⁻² within 12 months. This vegetation recovery drove ET enhancement from 0.35 to 2.65 mm·day⁻¹, achieving 58.9% restoration by December 2023 (Table 1). Notably, ET response lagged behind LAI increase by approximately 1–2 months, suggesting hydraulic system reconstruction requirements.

Stage III (stabilization phase, March 2024–August 2025) showed continued but decelerated recovery, with LAI reaching 5.42 m²·m⁻² (78% of pre-fire) and ET attaining 4.65 mm·day⁻¹ (100% restoration ratio) by August 2025 (Table 1; Figure 3). The seasonal ET pattern became increasingly pronounced during this phase, with summer peaks (July 2024: 3.75 mm·day⁻¹; July 2025: 4.35 mm·day⁻¹) and winter minima (December 2024: 3.95 mm·day⁻¹), reflecting reestablished canopy transpiration dominance (Figure 2).

3.2. AI-Based ET Modeling Performance

The Taylor diagram evaluation revealed substantial performance differences among the four modeling approaches (Figure 4). The Coupled LSTM–Transformer ensemble model achieved optimal simulation accuracy with RMSE = 0.10 mm·day⁻¹, NSE = 0.98, and KGE = 0.98, significantly outperforming single-architecture models (Table 1).

Stage-specific performance analysis indicated model-dependent advantages across recovery phases. During Stage I, the Transformer model exhibited superior performance (RMSE = 0.07, NSE = 0.95) due to its global attention mechanism capturing long-range dependencies under sparse vegetation conditions. In contrast, the LSTM model showed advantages in Stage II (rapid LAI transition period) with better temporal memory for lagged ET responses. The ensemble model consistently minimized errors across all stages, with maximum accuracy in Stage III (RMSE = 0.10, NSE = 0.97) when vegetation structure stabilized (Table 1).

The time-series comparison demonstrated that all AI models (LSTM, Transformer, Coupled) tracked observed ET dynamics more closely than the physical PML baseline, particularly capturing the nonlinear ET surge during Stage II transition (March–July 2023) when LAI increased from 0.35 to 0.92 m²·m⁻² (Figure 5; Table 1). The Coupled model reduced prediction uncertainty by 31% compared to LSTM alone, highlighting the complementary strengths of local temporal memory (LSTM) and global pattern recognition (Transformer).

SHAP analysis revealed phase-dependent shifts in dominant controlling factors (

Table 2. Phase-dependent evolution of SHAP feature importance identifying dominant ET drivers.

CF	Stage I	Stage II	Stage III	Average	Plot A	Plot B
LAI	15.2	52.8	48.5	45.2	48.5	41.8
VPD	28.5	22.5	26.2	25.1	23.8	26.5
SWC	42.5	12.5	5.5	14.8	12.5	17.2
PAR	9.5	8.2	15.8	9.8	9.5	10.2
Ta	3	3.2	3.2	3.2	3.8	2.5
WS	0.8	0.8	0.8	1.5	1.2	1.8
PRCP	0.5	0	0	0.4	0.7	0

; Figure 5). During Stage I, soil water content (SWC) dominated ET variation (42.5% contribution), followed by vapor pressure deficit (VPD, 28.5%), while LAI contributed minimally (15.2%) due to negligible canopy cover. This pattern inverted dramatically in Stage II, with LAI becoming the primary driver (52.8% contribution) as vegetation reestablished, while SWC contribution declined to 12.5%. Stage III maintained LAI dominance (48.5%) but showed increased photosynthetically active radiation (PAR) importance (15.8%) as canopy closure enhanced light competition (Table 2).

The LAI time-lag effect on ET exhibited a characteristic bimodal pattern (Figure 6;

Table 3. Burn severity effects on LAI-ET time-lag characteristics and ecological mechanisms. Time-lag peaks indicate days required for LAI changes to maximally influence ET.

Burn Severity	First Peak	Second Peak	Time-Lag Characteristic	Ecological Mechanism
Unburned	5 days	12 days	Short lag period	Healthy vegetation with an intact hydraulic system, rapid response
Low Severity	6 days	13 days	Similar to unburned	Minor damage, fast recovery
Moderate Severity	8 days	16 days	Significantly extended	Partial root damage, impaired water conduction
High Severity	12 days	21 days	Substantially delayed	Severe root destruction, hydraulic system reconstruction required

). The first peak occurred at 4–7 days (SHAP = 0.125, 27.7% contribution), representing leaf water potential equilibrium following LAI changes. The second peak emerged at 8–14 days (SHAP = 0.098, 21.7% contribution), corresponding to root water uptake compensation. Cumulative contributions reached 88.6% by day 14, indicating that LAI effects on ET are predominantly realized within two weeks.

Burn severity significantly modulated time-lag characteristics (Table 3). Unburned controls showed minimal lag (first peak: 5 days; second peak: 12 days), reflecting intact hydraulic systems. Lag periods extended progressively with burn severity: light severity (6/13 days), moderate severity (8/16 days), and severe burns (12/21 days)—more than doubling the unburned baseline. This severity-dependent extension indicates progressive damage to root systems and hydraulic architecture, requiring substantial reconstruction before efficient LAI-ET coupling reestablishes.

The ET-LAI response curves demonstrated model-specific physiological realism (Figure 6). All models captured the saturating logarithmic relationship (ET = a·ln(LAI + b) + c), but the Coupled model achieved the highest R² (0.98) and closest alignment with theoretical expectations (saturation LAI = 7.5, maximum ET = 6.2 mm·day⁻¹). Notably, severe burn plots exhibited 12%–15% lower ET than unburned controls at equivalent LAI (5.0 m²·m⁻²: 4.45 vs. 5.85 mm·day⁻¹ in Jinyun), suggesting persistent hydraulic limitation even after visual canopy recovery (Figure 7).

4. Discussion

4.1. Coupled AI Model Outperforms Single Architectures by Capturing Phase-Dependent Ecohydrological Dynamics

The coupled LSTM–Transformer ensemble achieved superior ET prediction accuracy (NSE = 0.98), reducing uncertainty by 31% compared to single models. This performance reflects complementary strengths: LSTM preserves local temporal memory while Transformer captures global pattern recognition, together resolving post-fire ecohydrological complexity more effectively than either architecture alone [20,21]. Physics-constrained loss functions, penalizing water balance violations and theoretical ET-LAI inconsistencies, ensured mechanistic realism without sacrificing predictive power—a critical advance over purely data-driven approaches [22]. Phase-specific model advantages revealed distinct regimes: Transformer superiority in Stage I (sparse vegetation) reflects global attention capturing atmospheric controls when canopy influence is minimal; LSTM advantages in Stage II indicate temporal memory importance for lagged vegetation–water coupling during structural reorganization. These findings suggest adaptive model selection or dynamic ensemble weighting could further enhance predictive capability [23].

4.2. Phase-Dependent Factor Transitions and Hydraulic Time-Lag Effects Reveal Post-Fire Recovery Mechanisms

SHAP analysis identified fundamental shifts in ET controls: soil water content dominated Stage I (42.5%), resembling bare soil evaporation regimes [9], while LAI controlled Stages II–III (>48%), indicating threshold behavior where canopy becomes the primary water-atmosphere exchange interface [24]. Radiation importance increased to 15.8% in Stage III, suggesting light limitation emerges as the canopy saturates [25]. Critically, we discovered characteristic bimodal LAI-ET time-lags: 4–7 days (leaf water potential equilibrium, 27.7%) and 8–14 days (root compensation, 21.7%), with 88.6% contribution within 14 days. The first peak reflects hydraulic capacitance—new leaves require days to establish functional transport from stem storage [26]. The second peak corresponds to root system activation supporting expanded canopy demand. Severity-dependent lag extension—from 5/12 days (unburned) to 12/21 days (severe)—directly quantifies fire damage to hydraulic architecture, requiring fine-root and mycorrhizal network reconstruction before efficient coupling reestablishes [27,28].

4.3. Persistent Hydraulic Limitations and Management Implications for Post-Fire Recovery Assessment

The 12%–15% ET reduction in severe burns at equivalent LAI reveals “hidden” hydraulic dysfunction not apparent from structural metrics alone, suggesting fire-induced xylem damage or root simplification reduces transport efficiency per unit leaf area [29,30]. Such limitations may persist years beyond visual recovery, with implications for carbon sequestration, streamflow generation, and drought vulnerability [31,32]. Management applications include: (1) extending recovery assessment timelines beyond canopy metrics to include functional ET monitoring; (2) prioritizing soil and root system protection in post-fire rehabilitation; and (3) incorporating AI-ET models into early warning systems for fire risk assessment. The severity-dependent time-lag parameters provide quantitative recovery targets: sites failing to achieve a <10-day LAI-ET response within 2 years likely require intervention. Future work should integrate drone-based LiDAR for spatial scaling [33] and test model transferability across forest types and climate zones, with continued monitoring to validate long-term projections. The physics-constrained framework facilitates such transfer through adjustable parameters and retrainable AI components, enabling integration with process-based vegetation models [8].

5. Conclusions

This study advances post-fire ecohydrology through four key contributions. First, physics-constrained ensemble AI (Coupled LSTM–Transformer) achieves superior ET prediction accuracy (NSE = 0.98) while maintaining mechanistic interpretability, outperforming single models by 31% in uncertainty reduction. Second, SHAP analysis reveals phase-dependent factor shifts from soil moisture dominance (Stage I, 42.5%) to LAI control (Stages II–III, >48%), with radiation limitation emerging in mature recovery, providing quantitative targets for monitoring design. Third, characteristic bimodal LAI-ET time-lag effects (4–7 and 8–14 days) extend progressively with burn severity (from 5/12 days unburned to 12/21 days severe), directly quantifying hydraulic system reconstruction requirements. Fourth, ET-LAI response curves reveal a persistent 12%–15% ET reduction in severe burns despite equivalent LAI recovery, indicating lasting hydraulic dysfunction invisible to structural assessment alone. The 2022 Chongqing wildfires exemplify climate change-driven fire regime expansion; our findings demonstrate that recovery assessment must extend beyond visual canopy metrics to include functional ET monitoring and time-lag characterization. The AI-SHAP framework provides an interpretable ecohydrological modeling template applicable to diverse disturbance contexts globally.