Plant phenology modelling aims to predict the timing of seasonal phases, such as leaf-out or flowering, from meteorological time series. Reliable predictions are crucial for anticipating ecosystem responses to climate change. While phenology modelling has traditionally relied on mechanistic approaches, deep learning methods have recently been proposed as flexible, data-driven alternatives with often superior performance. However, mechanistic models tend to outperform deep networks when data distribution shifts are induced by climate change. Domain Adaptation (DA) techniques could help address this limitation. Yet, unlike standard DA settings, climate change induces a temporal continuum of domains and involves both a covariate and label shift, with warmer records and earlier start of spring. To tackle this challenge, we introduce Mid-feature Rank-adversarial Domain Adaptation (MIRANDA). Whereas conventional adversarial methods enforce domain invariance on final latent representations, an approach that does not explicitly address label shift, we apply adversarial regularization to intermediate features. Moreover, instead of a binary domain-classification objective, we employ a rank-based objective that enforces year-invariance in the learned meteorological representations. On a country-scale dataset spanning 70 years and comprising 67,800 phenological observations of 5 tree species, we demonstrate that, unlike conventional DA approaches, MIRANDA improves robustness to climatic distribution shifts and narrows the performance gap with mechanistic models.
Figures & tables
Figure 1: Climate change-robust machine learning. Deep learning models trained on historical climatic conditions face a distribution shift when applied to future climate projections. In this paper we consider models that predict plant phenological dates from meteorological time series and devise a domain adaptation method to enhance their robustness to such shifts. (Temperature data: CMIP6 CNRM-CM6-1 SSP2-4.5 [ 12 , 38 ] , phenological data: Swiss Phenology Network [ 13 ] )
Figure 2: Phenology modeling (conceptual illustration). A phenology model takes a set of climatic and environmental covariates and predicts phenological dates. In this paper we consider models operating on daily meteorological time series to predict the date of 50% leaf development of 5 different tree species.
Figure 3: Method overview . Our framework addresses domain shifts in phenology modelling via two key components: rank-based adversarial training on intermediate features and hybrid layer normalization. Purple elements correspond to the main phenology prediction pathway : the learnable tokens L are concatenated with the input time series Ei and processed by two transformer encoder layers ( t1 and t2 ). The resulting global embeddings of the learnable tokens, denoted as Gi , are then fed into the decoder d to predict the target date using the regression loss LMSE . We apply rank-based adversarial learning on the mid-level features Zi (green), where a discriminator p is trained with a ranking loss Lrank through a gradient reversal layer to encourage domain-invariant mid-level features. Meanwhile, we replace the standard layer normalization in t2 with our hybrid layer normalization (blue), which preserves domain-dependent variations in high-level representations.
Zs=t1(xs),Zt=t1(xt)
Algorithm 1 Overall training framework
Figure 4: Domain distribution shifts. We show the histograms of input (top) and predicted (bottom) variables on the train, validation, and test sets for each of the three dataset splits. We plot the histogram of annual temperature in the top row to represent the climatic distribution, and the date of Larch needle emergence in the bottom row to illustrate the shift in predicted variables.
Dataset split →
1. Chronological
2. Annual Temperature
3. Elevation
Training
Model
R2↑
RMSE ↓
MAE ↓
R2↑
RMSE ↓
MAE ↓
R2↑
RMSE ↓
MAE ↓
time (min)
PhenoFormer Vanilla [ 15 ]
0.50 (0.01)
8.96
6.82
0.41 (0.02)
9.56
7.26
0.17 (0.02)
11.67
9.00
15
+DANN [ 1 ]
0.49 (0.01)
9.08
6.98
0.38 (0.06)
9.78
7.44
-0.20 (0.21)
14.03
11.32
17
+ADDA [ 37 ]
0.23 (0.15)
11.14
8.80
0.12 (0.10)
11.64
9.18
-0.61 (0.16)
16.22
13.33
26
+CORAL [ 35 , 34 ]
0.51 (0.01)
8.93
6.79
0.42 (0.02)
9.50
7.22
-0.16 (0.15)
13.77
10.98
18
+DANL [ 30 ]
0.48 (0.02)
9.17
7.03
0.30 (0.03)
10.44
8.06
-0.01 (0.10)
12.90
10.24
18
Table 1: Main results of three data splits. We compare MIRANDA with PhenoFormer and five domain adaptation baselines across three data splits representing progressively stronger distribution shifts: Chronological, Annual Temperature, and Elevation. For reference, we also report the performance of the process-based model (M1), highlighted in blue. Evaluation metrics include R2 ( ↑ ), RMSE ( ↓ ), and MAE ( ↓ ). Results are averaged over 10 repetitions and 5 tree species, and we also report the standard deviation of R2 in brackets.
Figure 5: Per-species performance Histogram of R2 per species for PhenoFormer and our model. Five species are: horse chestnut (HCH), European beech (BEE), European larch (LAR), common spruce (SPR), and hazel (HZL).
Figure 6: Predicted (y-axis) vs true dates (x-axis) before (left, with R2 = 17%) and after (right, with R2 = 22%) adaptation with MIRANDA on the elevation split. Note that we plot the dates of the 5 species together, which explains the different modes of the distribution.
Dataset split →
1. Chronological
2. Annual Temperature
3. Elevation
R2↑
RMSE ↓
MAE ↓
R2↑
RMSE ↓
MAE ↓
R2↑
RMSE ↓
MAE ↓
PhenoFormer Vanilla
0.50
8.96
6.82
0.41
9.56
7.26
0.17
11.67
9.00
+DANN [ 1 ]
0.49
9.08
6.98
0.38
9.78
7.44
-0.20
14.03
11.32
✓mid-feat
0.51
8.96
6.82
0.40
9.59
7.23
0.12
12.04
9.38
✓mid-feat ✓ Lrank
0.50
8.97
6.82
0.42
9.45
7.18
0.19
11.54
8.87
✓mid-feat ✓ hLN
0.50
8.97
6.81
0.42
9.51
7.22
0.18
11.66
8.97
Table 2: Ablations study of MIRANDA . Starting from the PhenoFormer backbone and DANN adversarial training, we progressively introduce (1) mid-level feature alignment (mid-feat), (2) rank-based adversarial objective Lrank , and (3) hybrid layer normalization ( hLN ). Results demonstrate that each component contributes to improved robustness under domain shift, with the largest gains observed under the challenging Elevation split.
Guidance
R2↑
RMSE ↓
MAE ↓
Annual temperature
0.50
9.01
6.85
Elevation
0.50
9.01
6.89
Year (ours)
0.51
8.97
6.80
Table 3: Different guidance for rank-based adversarial training under the Structured Temporal split.
Guidance
R2↑
RMSE ↓
MAE ↓
Annual temperature
0.43
9.44
7.18
Elevation
0.42
9.44
7.19
Year (ours)
0.43
9.43
7.16
Table 4: Different guidance for rank-based adversarial training under the Annual Temperature split.
Guidance
R2↑
RMSE ↓
MAE ↓
Annual temperature
0.22
11.36
8.67
Elevation
0.19
11.55
8.86
Year (ours)
0.22
11.32
8.63
Table 5: Different guidance for rank-based adversarial training under the Elevation split.
Figure 1: Scatter plot of predicted (y-axis) vs true dates (x-axis) before (left, with R2 = 24%) and after (right, with R2 = 26%) adaptation for phenology task leaf unholding of horse chestnut in elevation data split.
Figure 2: Scatter plot of predicted (y-axis) vs true dates (x-axis) before (left, with R2 = 12%) and after (right, with R2 = 23%) adaptation for phenology task needle emergence of European larch in elevation data split.
Figure 3: Scatter plot of predicted (y-axis) vs true dates (x-axis) before (left, with R2 = 12%) and after (right, with R2 = 15%) adaptation for phenology task leaf unholding of hazel in elevation data split.
Figure 4: Scatter plot of predicted (y-axis) vs true dates (x-axis) before (left, with R2 = 25%) and after (right, with R2 = 28%) adaptation for phenology task needle emergence of common spruce in elevation data split.
Figure 5: Scatter plot of predicted (y-axis) vs true dates (x-axis) before (left, with R2 = 13%) and after (right, with R2 = 20%) adaptation for phenology task leaf unfolding of European beech in elevation data split.
Sustainability and Data Sciences Laboratory, Northeastern University · Microsoft · AI for Climate and Sustainability, The Institute for Experiential AI, Northeastern University