Source-linked AI summary
Learning Regional Snow Water Equivalent and Snow Height Variations from Sentinel-1 InSAR Acquisitions
Luca Barco, Lorenzo Innocenti, Bianca Bartoli, Claudio Rossi, Edoardo Arnaudo, Paolo Garza
TL;DR
Reliable SWE and HS data are important for mountainous water management, but large-scale monitoring remains difficult. This study compares XGBoost, U-Net, and SegFormer for joint estimation of SWE and HS variations from Sentinel-1 InSAR data against IT-SNOW; SegFormer has the lowest errors and variability, while spatial metrics distinguish architectures more clearly than mean errors.
Problem
Reliable SWE and HS measurements are needed for mountainous water management, but sparse ground observations limit snow monitoring across large alpine areas.
Source: §I · p. 1Method
The study compares XGBoost, U-Net, and SegFormer for joint prediction of SWE and HS variations from Sentinel-1 InSAR data using IT-SNOW as reference.
Source: §I · p. 2, B. Model Architectures · p. 3Results
SegFormer has the lowest error and variability on both targets; on the all-features baseline, its MAE is 10.391 cm for HS and 27.113 mm w.e. for SWE.
Source: §V · p. 4, B. Results · p. 4Takeaways & Limitations
Spatial R2 and Pearson’s r separate the three architectures more clearly than mean error, and multi-target training consistently improves accuracy.
Source: §V · p. 4Across seven test windows, the study cannot establish systematic dependence on thermal regime or estimate the window-mean offset independently for calibration.
Source: §V · p. 5
Abstract
from arXiv · showhide
Managing water resources in mountainous regions depends heavily on reliable Snow Water Equivalent (SWE) and Snow Height (HS) data, yet these variables remain difficult to track at scale. This study evaluates three machine learning architectures (XGBoost, U-Net and SegFormer) for the joint estimation of SWE and HS variations from Sentinel-1 InSAR data over the Italian Alps, using the IT-SNOW reanalysis as reference. SegFormer achieves the best results on both targets, with an MAE of 10.391 cm for HS and 27.113 mm w.e. for SWE and the lowest variability across initializations. A feature sensitivity analysis shows that including all available features does not guarantee the lowest error, with model- and task-specific sensitivities. Spatial metrics (R2, Pearson's r) separate the three architectures far more clearly than mean error (MAE, RMSE) does, and decomposing the error per window attributes most of it to a systematic offset in the estimated mean variation rather than to the spatial pattern.
I. INTRODUCTION
The introduction motivates scalable estimation of SWE and HS for mountainous water management and identifies gaps in joint retrieval, feature selection, and architecture comparison. It compares three ML approaches for Sentinel-1 InSAR data over the Italian Alps, with SegFormer reported to achieve the lowest error and variability.
I. INTRODUCTION
- Ground measurements are accurate but sparse, while Sentinel-1 SAR offers a scalable, all-weather option for monitoring snow over large alpine domains.InSAR estimates SWE variations but can be limited by temporal decorrelation; amplitude-based HS transferability across C-band settings remains under investigation.Source: §I · p. 1
- Prior work leaves open how to estimate SWE and HS jointly, choose InSAR features, and compare ML paradigms using the same aggregated observables.The cited approaches primarily estimate single targets or absolute snow quantities over smaller domains.Source: §I · p. 1
- The study compares XGBoost, U-Net with a ResNet backbone, and SegFormer for simultaneous SWE and HS variation retrieval from Sentinel-1 InSAR data.Its contributions include feature-sensitivity analysis and group-wise ablation of interferometric channels.Source: §I · p. 1
- SegFormer attains the lowest error and variability across initializations, while spatial correlation metrics separate the architectures more clearly than mean error.Training covers 09 January 2021–30 December 2022, with an independent test period through 24 March 2023.Source: §I · p. 2
II. DATASET
The dataset covers the Orobie Alps and Adamello and combines Sentinel-1 InSAR features with thermal, geometric, and temporal inputs. IT-SNOW supplies 500 m reference data, from which SWE and HS changes between acquisitions are derived.
II. DATASET
- The case study covers the Orobie Alps and Adamello, whose varied terrain and snow accumulation patterns represent complex alpine environments.Source: §II · p. 2
- Sentinel-1A SLC images in IW VV mode from ascending relative orbit 117, with a 12-day repeat cycle, were processed in SNAP2 to produce interferograms, coherence estimates, phase, and displacement maps.Source: §II · p. 2
- The inputs include five interferometric features, two geometric variables, mean land-surface temperature, positive degree days, and four date encodings.The geometric inputs are local incidence angle and elevation; the date features encode acquisition start and end times.Source: §II · p. 2
- IT-SNOW provides daily 500 m HS and SWE reference values, and labels are differences between consecutive SAR acquisition dates across 64 temporal samples.All sources were bilinearly resampled to 500 m, and areas outside Italian borders were excluded.Source: §II · p. 2
III. METHODOLOGY
The study compares three machine-learning paradigms for SWE and HS estimation: gradient boosting, a convolutional encoder–decoder, and a vision transformer.
III. METHODOLOGY
- The comparative analysis evaluates XGBoost, U-Net with a ResNet backbone, and SegFormer.Source: §III · p. 2
- XGBoost represents classical gradient boosting in the architecture comparison.Source: §III · p. 2
- U-Net and SegFormer represent convolutional encoder–decoder and vision-transformer paradigms, respectively.Source: §III · p. 2
A. Problem Formulation
The task is supervised multi-target regression of snow changes between SAR acquisitions. InSAR-derived input features are used to predict spatially distributed SWE and HS change maps.
A. Problem Formulation
- The formulation treats SWE and HS estimation as supervised regression of temporal changes rather than daily absolute values.Source: A. Problem Formulation · p. 2
- ΔSWE and ΔHS are defined as the values at t2 minus those at t1, the reference and secondary SAR acquisition dates.Source: A. Problem Formulation · p. 2
- A spatial input tensor of InSAR, LST, and temporal features is used to predict spatially distributed snow-change maps.The input has dimensions H × W × C, and the output map has dimensions H × W.Source: A. Problem Formulation · p. 2
B. Model Architectures
The section compares three architectures for snow-map prediction: pixel-wise gradient boosting, convolutional encoder–decoder learning, and a Transformer-based model. U-Net and SegFormer produce joint HS and SWE maps, while performance is evaluated with four error and spatial metrics.
B. Model Architectures
- XGBoost builds sequential decision-tree ensembles and treats each pixel independently, without explicitly using neighboring-pixel relationships.Separate XGBoost models are used for HS and SWE.Source: B. Model Architectures · p. 3
- U-Net combines a contracting encoder and expansive decoder with a ResNet-18 backbone to produce joint HS and SWE maps.Skip connections combine encoder features with upsampled maps as spatial resolution is recovered.Source: B. Model Architectures · p. 3
- SegFormer uses a hierarchical MiT-B0 Transformer encoder and lightweight MLP decoder to aggregate multi-scale features into HS and SWE maps.Its self-attention operates on image patches to capture long-range dependencies and global context.Source: B. Model Architectures · p. 3
- Performance is evaluated with MAE, RMSE, R2, and Pearson’s correlation coefficient r.Source: B. Model Architectures · p. 3
A. Implementation details
Experiments use temporally separated training, validation, and independent test windows, with three random initializations. The section also describes model-specific input handling and one-feature-at-a-time sensitivity tests.
A. Implementation details
- Training used 48 samples, validation used 9 temporally spaced samples, and the final 7 windows formed an independent test set; each experiment used three random initializations.The training period ended 30 December 2022, and the held-out test period ran through 24 March 2023.Source: A. Implementation details · p. 3
- XGBoost treats pixels independently with separate HS and SWE models and native missing-value handling, whereas U-Net and SegFormer use joint patch regression with zero imputation.The neural models share 256 × 256 patches, AdamW, cosine annealing, and masked Huber loss.Source: A. Implementation details · p. 3
- Sensitivity tests excluded each feature in turn on the test set while retaining thermal and temporal encodings; Tables I–III report results for XGBoost, U-Net, and SegFormer.Source: A. Implementation details · p. 3, Table I · p. 3, Table II · p. 3
B. Results
SegFormer’s spatial metrics outperform XGBoost’s with the complete feature set, while error decomposition attributes most normalized error to window-mean offsets. Feature effects vary by model and target, and removing InSAR bands degrades SegFormer’s metrics.
B. Results
- Excluding features can reduce error, but the best subset depends on model and target; SegFormer’s HS error is lowest without unwrapped phase, but displacement exclusion is more stable.For SegFormer, excluding unwrapped phase gives 9.799 cm mean error with 3.749 cm standard deviation; excluding displacement gives 9.856 cm with 0.783 cm standard deviation.Source: B. Results · p. 3
- Every single-feature exclusion improves XGBoost over its all-features configuration; its HS MAE falls from 14.429 to 12.374–13.389 cm, yet it remains less accurate than SegFormer.The full configuration is also least stable across seeds for HS RMSE, and the best exclusion differs by target.Source: B. Results · pp. 3–4
- R2 is −0.800 for SegFormer, versus −2.138 for U-Net and −2.631 for XGBoost; Pearson’s r is 0.065, 0.018, and 0.002, respectively, with complete features.Table IV covers spatial metrics for the three architectures and compares multi-target with single-target training.Source: B. Results · p. 4, Table IV · p. 4
- 58–73% of normalized error comes from systematic offsets in window-mean variation; the remainder reflects underdispersion and near-zero pattern correlation.All models underpredict reference variability, and the low r values indicate that the spatial signal is largely unrecovered.Source: B. Results · p. 4
- Removing SegFormer’s five interferometric bands raises MAE from 10.391 to 11.082 cm for HS and from 27.113 to 28.143 mm w.e. for SWE.The ablation also worsens R2, Pearson’s r, predicted-to-reference variability ratios, and seed-to-seed MAE spread.Source: B. Results · p. 4
- For one test pair, SegFormer’s HS map is closest to the reference, while XGBoost’s SWE r of 0.55 largely reflects the elevation gradient.On HS, XGBoost’s MAE is 15.13 cm against 8.88 cm for SegFormer; the single-window comparison does not reliably separate architectures overall.Source: B. Results · p. 4
V. CONCLUSIONS AND FUTURE WORKS
SegFormer has the lowest error and variability on both snow-change targets, while spatial metrics distinguish architectures more clearly than mean error. The conclusions also identify limits from low coherence, short test coverage, reference validation, and 500 m aggregation.
V. CONCLUSIONS AND FUTURE WORKS
- SegFormer attains the lowest error and variability on both targets; spatial metrics separate architectures more clearly than mean error, and all-feature inputs do not ensure lowest error.The informative feature subset is model- and task-specific.Source: §V · p. 4
- Multi-target training consistently improves accuracy, while removing interferometric channels degrades every metric.Source: §V · p. 4
- Mean coherence remains 0.26–0.33, constraining phase-feature contributions; across seven test windows, spatial correlation varies without an established thermal-regime dependence.Source: §V · pp. 4–5
- A systematic offset accounts for 58–73% of normalized error, but the test period cannot estimate it independently or separate it from the training set’s seasonal composition.All models underpredict reference variability; the offset is described as a calibration target, not an intrinsic limit.Source: §V · p. 5
- The test-period reference lacks published validation, independent in-situ validation was unavailable, and 500 m aggregation averages out sub-pixel variability in complex terrain.Source: §V · p. 5