- Rating
- 2(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
The authors introduce an uncertainty-aware temporal distillation framework using precomputed annual geospatial foundation model embeddings as supervision to predict in-season embeddings from monthly Sentinel-2 observations. The framework is evaluated on how well monthly student embeddings approximate annual teacher embeddings from the same year, as well as on downstream cropland classification, change detection, and field boundary segmentation. These tasks demonstrate that the predicted embeddings approach the annual teacher representations and achieve competitive or better downstream performance later in the year (with the most temporal context).
Critical strengths
• The introduction clearly motivates the need for in-season, on-demand representations without recomputing expensive geospatial foundation model embedding products.
• The method shows the ability to reconstruct annual embeddings from monthly Sentinel-2 imagery and perform well at downstream tasks for months with more temporal context (later months of the year).
Critical weaknesses
• It is unclear why approximating a calendar-year teacher embedding is useful for in-season crop monitoring. This is especially clear in the poor downstream performance for the early months of the year, which seems to conflict with the motivation of providing timely, in-season representations.
• The authors do not provide evidence that the distillation objective provides an advantage over simpler alternatives for in-season prediction. There is no comparison against a similarly sized temporal encoder trained using self-supervised or seasonal prediction objectives without annual teacher supervision.
• The motivation for monthly embeddings is not sufficiently justified. Downstream performance tasks are on a seasonal scale (even more coarse than monthly) and it is plausible that coarser or finer temporal resolution would be useful for different downstream tasks
• The strongest downstream classification results come from accumulated monthly student embeddings rather than the single embedding from the month of interest. It is unclear whether the improved downstream performance is due to the increased feature dimensionality or whether similar improvements could be achieved with simpler aggregation techniques (e.g., mean pooling of embeddings).
• The discussion of existing geospatial foundation models and embedding products could more clearly distinguish between the temporal limitations of precomputed embedding products and the capabilities of the foundation models. Models like Clay, Prithvi-EO, and OlmoEarth can generate representations from observations at flexible temporal resolutions
• Figure captions / titles need more explanation (minor: why are the RGB images in figure 2 discolored)
Questions for authors
• Why monthly temporal resolution? How would performance change with seasonal composites or phenology-aligned observations? How temporally granular can you go, how do different temporal resolutions affect stability and performance at different levels of temporal aggregation (how sensitive are the predictions to cloud cover and missing observations)?
• Could the authors clarify how monthly predictions are constructed across calendar-year boundaries, particularly whether observations from the preceding autumn are retained or the input sequence resets in January (Figure 4 Ukraine)?
• Have the authors considered using a rolling 12-month observation window rather than resetting the input sequence each calendar year? This could potentially provide more consistent temporal context (and potentially more consistent performance throughout the year, which fits well with the motivation of the work)
• Since the distillation dataset is from 2024, how well does this generalize to years with different weather or agricultural patterns? How does performance change year-to-year on the downstream datasets with points from other years?
• How does downstream performance compare to a similarly sized temporal encoder trained with a seasonal prediction or self-supervised objective without annual embedding distillation?
• Could the authors provide anonymized access to the code, inference pipeline, and trained models during review through the supplementary material and/or anonymized links?
15 comments
Responses 2/4
[W4] The strongest downstream classification results come from accumulated monthly student embeddings rather than the single embedding from the month of interest. It is unclear whether the improved downstream performance is due to the increased feature dimensionality or whether similar improvements could be achieved with simpler aggregation techniques (e.g., mean pooling of embeddings).
We show that our student outperforms the oracle in Sudan and is competitive in Ukraine at matched feature dimensionality by using simple aggregation techniques. Please note that the current-year AEF (oracle) only becomes available after the full year is observed, so it’s a fair baseline for the full calendar-year comparison here but not for other sub-calendar-year comparisons. Table Note: The best F1 Score is highlighted in bold and the second-best in italics.
| Features | Dim. | Sudan F1 ↑ | Ukraine F1 ↑ |
|---|---|---|---|
| Full calendar-year: Monthly embedding aggregation | |||
| Mean pooling | 64 | 0.7976 | 0.8808 |
| Max pooling | 64 | 0.8016 | 0.8718 |
| Last month only | 64 | 0.7981 | 0.9015 |
| All monthly embeddings (PCA) | 64 | 0.8259 | 0.8993 |
| Reference baselines | |||
| Previous-year AEF | 64 | 0.7963 | 0.8584 |
| Annual AEF (oracle) | 64 | 0.8133 | 0.9173 |
[W5] The discussion of existing geospatial foundation models and embedding products could more clearly distinguish between the temporal limitations of precomputed embedding products and the capabilities of the foundation models. Models like Clay, Prithvi-EO, and OlmoEarth can generate representations from observations at flexible temporal resolutions
TLDR: We show that our student model performs competitively with monthly embeddings from TESERRA and OlmoEarth-Nano while being more efficient.
Global embeddings products like AEF, OlmoEarth, and TESERRA do not publish continuous in-season updates, only annual ones. AEF’s monthly embeddings are only available in private preview, so we cannot compare with them. We run TESERRA to get monthly embeddings and show that our student model performs competitively (outperforms early in the calendar-year) with TESERRA while being 50x smaller and 4.8x higher in throughput (). Similarly, in the non-pixel time-series GeoFMs family, we run the similarly sized OlmoEarth-Nano monthly and compare it to our student model. We observed similar performance trends, while our model has 1.3x higher throughput (). This shows that our model is a more efficient alternative for frequent in-season monitoring.
Sudan
| Cutoff | TESSERA F1 ↑ | OlmoEarth Nano F1 ↑ | AEF student F1 ↑ |
|---|---|---|---|
| Jan | 0.7382 | 0.7353 | 0.7958 |
| Feb | 0.7314 | 0.7459 | 0.7940 |
| Mar | 0.7503 | 0.7506 | 0.7918 |
| Apr | 0.7576 | 0.7562 | 0.7922 |
| May | 0.7627 | 0.7671 | 0.7932 |
| Jun | 0.7631 | 0.7823 | 0.7952 |
| Jul | 0.7933 | 0.7972 | 0.8047 |
| Aug | 0.8116 | 0.8021 | 0.7996 |
| Sep | 0.8239 | 0.8166 | 0.8009 |
| Oct | 0.8227 | 0.8327 | 0.8032 |
| Nov | 0.8360 | 0.8391 | 0.8021 |
| Dec | 0.8411 | 0.8418 | 0.7981 |
Ukraine
| Cutoff | TESSERA F1 ↑ | OlmoEarth Nano F1 ↑ | AEF student F1 ↑ |
|---|---|---|---|
| Sep′ | 0.8602 | 0.8085 | 0.8432 |
| Oct′ | 0.8637 | 0.8315 | 0.8451 |
| Nov′ | 0.8663 | 0.8402 | 0.8402 |
| Dec′ | 0.8705 | 0.8416 | 0.8422 |
| Jan | 0.8046 | 0.8467 | 0.8625 |
| Feb | 0.8200 | 0.8467 | 0.8633 |
| Mar | 0.8370 | 0.8464 | 0.8673 |
| Apr | 0.8517 | 0.8494 | 0.8677 |
| May | 0.8782 | 0.8604 | 0.8742 |
| Jun | 0.8961 | 0.8662 | 0.8796 |
| Jul | 0.9219 | 0.8866 | 0.8904 |
| Aug | 0.9270 | 0.8960 | 0.8965 |
| Sep | 0.9324 | 0.8994 | 0.8997 |
| Oct | 0.9304 | 0.8974 | 0.9018 |
| Nov | 0.9337 | 0.8949 | 0.9035 |
| Dec | 0.9320 | 0.8956 | 0.9015 |
[W6] Figure captions / titles need more explanation (minor: why are the RGB images in figure 2 discolored)
Thank you for pointing this out. The PCA-RGB row in Figure 2 shows false colors, not true RGB. We will updated Figure 2 in the paper.
Responses 3/4
Responses to Questions
[Q1] Why monthly temporal resolution? How would performance change with seasonal composites or phenology-aligned observations? How temporally granular can you go, how do different temporal resolutions affect stability and performance at different levels of temporal aggregation (how sensitive are the predictions to cloud cover and missing observations)?
TLDR: i) Monthly temporal resolution is motivated by in-season monitoring products, ii) our methods can be adapted to different temporal resolution/cadence, and iii) our method shows an F1 score drop of only 0.0324 even with 50% missing observations.
This is related to your comment in W3 "The motivation for monthly embeddings is not sufficiently justified. Downstream performance tasks are on a seasonal scale (even more coarse than monthly), and it is plausible that coarser or finer temporal resolution would be useful for different downstream tasks". In our response, we showed that in-season monitoring products (Crop Monitor, Food Security Analysis, Food Assistance Outlook Briefs, and Global Information and Early Warning System (GIEWS) Food Price Monitoring) release reports monthly, which we aim to assist with our monthly temporal embeddings.
We show that our method can be adapted to any temporal cadence based on Sentinel-2 images. We also compare monthly embeddings to bimonthly and quarterly embeddings on our downstream tasks to show that monthly achieves the best performance.
Regarding robustness to missing observations; in our input data each month is a median composite of all Sentinel-2 scenes with less than 40% cloud cover. A month with no usable scene enters the model as an empty month. We tested the impact of missing months on the downstream performance and observed that with 6 of 12 monthly composites randomly dropped, F1 score only drops by 0.0324 in Sudan and 0.027 in Ukraine.
| Months removed | none | 1 random | 3 random | 6 random | 6 in a row |
|---|---|---|---|---|---|
| Sudan F1 | 0.8479 | 0.8442 | 0.8302 | 0.8155 | 0.8203 |
| Ukraine F1 | 0.9213 | 0.9156 | 0.9090 | 0.8943 | 0.8843 |
[Q2] Could the authors clarify how monthly predictions are constructed across calendar-year boundaries, particularly whether observations from the preceding autumn are retained or the input sequence resets in January (Figure 4 Ukraine)?
The raw input sequence resets in January. This does not cause an issue during inference, as we also provide full previous-year’s embeddings to the student. When a crop season crosses the calendar year boundaries, as in Figure 4 for Ukraine, the predictions for season Y come from two runs of the student.
| Cutoffs | Input months | Prior |
|---|---|---|
| Sep′ to Dec′ | the months of calendar year Y-1, up to the cutoff | AEF Y-2 |
| Jan to Dec | the months of calendar year Y, up to the cutoff | AEF Y-1 |
For example, we infer for Sep’ to Dec’ (say, 2023) by passing 2022 AEF and 2023 ground samples as inputs. Then we infer Jan to Nov (2024) by passing 2023 AEF and 2024 ground samples.
Q3 Have the authors considered using a rolling 12-month observation window rather than resetting the input sequence each calendar year? This could potentially provide more consistent temporal context (and potentially more consistent performance throughout the year, which fits well with the motivation of the work)
TLDR: Rolling 12-months provide minimal improvement compared to the calendar year.
We evaluated a rolling 12-month window against our calendar-year approach, using the same architecture, training settings, and previous-year AEF conditioning. The rolling model uses the most recent 12 months of Sentinel-2 observations, while the calendar model uses January through the current month. Both were trained with three seeds; we report the cosine similarity to the current-year annual AEF averaged over three seeds:
| Cutoff | 2024 calendar | 2024 rolling | 2025 calendar | 2025 rolling |
|---|---|---|---|---|
| Jan | 0.9461 | 0.9484 | 0.9430 | 0.9443 |
| Mar | 0.9501 | 0.9511 | 0.9467 | 0.9476 |
| Jun | 0.9557 | 0.9561 | 0.9540 | 0.9540 |
| Sep | 0.9612 | 0.9615 | 0.9607 | 0.9606 |
| Dec | 0.9621 | 0.9622 | 0.9615 | 0.9613 |
Rolling windows provide a small improvement in early-year cosine similarity, particularly in January (+0.0023 in 2024 and +0.0013 in 2025), but the advantage disappears by midyear. We also evaluated downstream cropland F1 in Sudan and Ukraine. The differences between calendar and rolling windows are small and inconsistent, with neither approach consistently outperforming the other.
Responses 1/4
Dear Reviewer gSpa,
We sincerely thank you for taking the time to review our paper and for your thoughtful comments and constructive suggestions. We have conducted additional experiments and analyses to address your concerns and provide responses below.
Responses to Critical Weaknesses
[W1] It is unclear why approximating a calendar-year teacher embedding is useful for in-season crop monitoring. This is especially clear in the poor downstream performance for the early months of the year, which seems to conflict with the motivation of providing timely, in-season representations.
In-season agricultural monitoring aims to support early prediction of end-of-season outcomes, such as crop yields and production, which are important inputs to food security early warning systems (Nakalembe & Kerner, 2023; Anderson et al., 2026). So the end-of-season or annual representation is exactly what users want, and that is what our student model learns to approximate.
Our student model is expected to perform poorly in early months compared to the current year AEF as the current year AEF has already seen the whole year while the student has only seen the early months. The fair baseline here is the previous year AEF which our student model outperforms or is comparable to. In Figure 4, we compare against this baseline. In January, the student achieves similar F1 to previous-year AEF in Sudan (0.7959 vs. 0.7963) and higher F1 in Ukraine (0.8686 vs. 0.8583). Performance improves as more observations become available, particularly where land use changes.
[W2] The authors do not provide evidence that the distillation objective provides an advantage over simpler alternatives for in-season prediction. There is no comparison against a similarly sized temporal encoder trained using self-supervised or seasonal prediction objectives without annual teacher supervision.
TLDR: Our student outperforms similar sized self-supervised temporal encoders and fully supervised similar architecture.
We tested two alternatives to annual teacher distillation:
1) Self-supervised temporal encoder. OlmoEarth Nano has 1.36M parameters, compared with 0.93M for our student. Using the same monthly observations, our student achieves higher early-season F1 (0.7958 vs. 0.7353 in Sudan in January). In December, Nano is competitive in Sudan (0.8412 vs. 0.8479), but remains below our student in Ukraine (0.9081 vs. 0.9213).
2) Full supervision. We trained our student architecture directly on cropland labels without annual teacher supervision. The distilled student with a linear probe performs better, including in December in Ukraine (0.9213 vs. 0.9103).
[W3] The motivation for monthly embeddings is not sufficiently justified. Downstream performance tasks are on a seasonal scale (even more coarse than monthly) and it is plausible that coarser or finer temporal resolution would be useful for different downstream tasks
TLDR: We show that our method can be adapted to different temporal resolution/cadence.
We use monthly observations because they match the update frequency of many operational agricultural monitoring products (examples below). Monthly composites also help reduce the impact of cloud cover while providing a practical balance between temporal detail and computational cost. Crop seasons also differ by location and crop type, making fixed seasonal intervals less general.
To evaluate coarser temporal resolutions, we used the same frozen model with bimonthly and quarterly sentinel-2 mosaics, without retraining. Changes are relative to monthly inputs, evaluated at the end of the calendar-year (December).