acceptodds
Under review as a conference paper at ICLR 2027

From Annual to In-Season: Uncertainty-Aware Distillation of Precomputed Geospatial Foundation Model Embeddings

Abstract

Geospatial foundation models (GFMs) have demonstrated strong performance across a wide range of downstream tasks. Globally precomputed GFM embeddings are increasingly being released as ready-to-use geospatial layers. However, these embeddings are typically annual and retrospective, limiting their use for within-year tasks. Generating new embeddings can also be computationally expensive and impractical at large scales. This precludes their use for temporally-sensitive tasks like agricultural monitoring, where embeddings must be available and re-computed frequently within the growing season. We present an uncertainty-aware framework that distills annual embeddings into temporally evolving representations using partial satellite time series observations. Our model predicts a distribution over the annual embedding conditioned on observations available at a given date. As the year progresses, predicted embeddings converge to the annual representation while uncertainty decreases. We show that the resulting embeddings support in-season agricultural monitoring tasks including cropland classification and field boundary segmentation. Student embeddings achieve high cosine similarity to the annual teacher embedding even with one month of data, and downstream performance matches or exceeds the annual teacher in most tasks. Our method offers a lightweight model to extend annual GFM embedding products to in-season agricultural applications.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Geospatial foundation models, knowledge distillation, precomputed embeddings, in-season monitoring, uncertainty estimation, Earth observation, agricultural monitoring, remote sensing

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_gSpaOpenReview ↗
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
AuthorsOpenReview ↗

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.

OrganizationOperational productReporting frequency
Global Agricultural Monitoring (GEOGLAM)Crop MonitorMonthly
Famine Early Warning Systems Network (FEWS NET)Food Security AnalysisMonthly
Famine Early Warning Systems Network (FEWS NET)Food Assistance Outlook BriefsMonthly
Food and Agriculture Organization of the United Nations (FAO)Global Information and Early Warning System (GIEWS) Food Price MonitoringMonthly

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).

Input frequencySudan Δ cosine ↑Ukraine Δ cosine ↑Sudan Δ F1 ↑Ukraine Δ F1 ↑
Bimonthly-0.0120-0.0088+0.0009-0.0071
Quarterly-0.0155-0.0157-0.0061-0.0144
AuthorsOpenReview ↗

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.

FeaturesDim.Sudan F1 ↑Ukraine F1 ↑
Full calendar-year: Monthly embedding aggregation
Mean pooling640.79760.8808
Max pooling640.80160.8718
Last month only640.79810.9015
All monthly embeddings (PCA)640.82590.8993
Reference baselines
Previous-year AEF640.79630.8584
Annual AEF (oracle)640.81330.9173

[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 (km2/skm^2/s). 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 (km2/skm^2/s). This shows that our model is a more efficient alternative for frequent in-season monitoring.

Sudan

CutoffTESSERA F1 ↑OlmoEarth Nano F1 ↑AEF student F1 ↑
Jan0.73820.73530.7958
Feb0.73140.74590.7940
Mar0.75030.75060.7918
Apr0.75760.75620.7922
May0.76270.76710.7932
Jun0.76310.78230.7952
Jul0.79330.79720.8047
Aug0.81160.80210.7996
Sep0.82390.81660.8009
Oct0.82270.83270.8032
Nov0.83600.83910.8021
Dec0.84110.84180.7981

Ukraine

CutoffTESSERA F1 ↑OlmoEarth Nano F1 ↑AEF student F1 ↑
Sep′0.86020.80850.8432
Oct′0.86370.83150.8451
Nov′0.86630.84020.8402
Dec′0.87050.84160.8422
Jan0.80460.84670.8625
Feb0.82000.84670.8633
Mar0.83700.84640.8673
Apr0.85170.84940.8677
May0.87820.86040.8742
Jun0.89610.86620.8796
Jul0.92190.88660.8904
Aug0.92700.89600.8965
Sep0.93240.89940.8997
Oct0.93040.89740.9018
Nov0.93370.89490.9035
Dec0.93200.89560.9015

[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.

AuthorsOpenReview ↗

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 removednone1 random3 random6 random6 in a row
Sudan F10.84790.84420.83020.81550.8203
Ukraine F10.92130.91560.90900.89430.8843

[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.

CutoffsInput monthsPrior
Sep′ to Dec′the months of calendar year Y-1, up to the cutoffAEF Y-2
Jan to Decthe months of calendar year Y, up to the cutoffAEF 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:

Cutoff2024 calendar2024 rolling2025 calendar2025 rolling
Jan0.94610.94840.94300.9443
Mar0.95010.95110.94670.9476
Jun0.95570.95610.95400.9540
Sep0.96120.96150.96070.9606
Dec0.96210.96220.96150.9613

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.

Reviewer_bDAPOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents a technique for creating geospatial embeddings for "with-in year tasks". Most geospatial foundation models only release the embeddings once the year is finished, as their model requires the entire sequence. The approach presented here uses the finished embedding as a supervision target and builds a model that uses last years embedding together with ongoing observations from each passed month to distill an in-season embedding with uncertainty. The results show that the student can retrieve the year-end embedding with more temporal observations and uncertainty decreases. Because of this in-season applications, like crop change detection are possible.

Critical strengths
  1. The paper is about a very important subject and presents a strong approach for tackling it. Having inter-annual embeddings without the need to apply the model multiple times or paying for platforms that do it for you (OlmoEarth) is a great contribution.
  2. The results show that the approach works as intended and allows to do intra-annual tasks that were not possible before (crop change detection).
  3. The paper is written well structured and understandable. Good job!
Critical weaknesses
  1. The only validation if intra-annual embeddings are actually useful is the crop change detection task. The other tasks are in my opinion not suited for testing whether the benefit comes from the intra-annual embedding or a small performance boost of the lightweight model.
Questions for authors
  1. Have you done an analysis about spatial autocorrelation or have you included a spatial buffer around your 15% validation dataset?
  2. Have you done ablations that test whether your last months embedding is actually better suited for other downstream tasks than the AEF embeddings? This would clear up if the benefit comes from the intra-annual embeddings or just a slightly better model
3 comments

Dear Reviewer bDAP,

We would like to thank you for constructive feedback and appreciating our writing! We have conducted additional analyses to address weaknesses and questions you identified and provide point-by-point responses below.

Responses to Critical Weaknesses

[W1] The only validation if intra-annual embeddings are actually useful is the crop change detection task. The other tasks are in my opinion not suited for testing whether the benefit comes from the intra-annual embedding or a small performance boost of the lightweight model.

We agree that crop change detection is the most direct way to evaluate the value of intra-annual embeddings, and our results demonstrate improvements on this task across Sudan and Ukraine. However, cropland classification and field boundary delineation also depend on seasonal information, as vegetation conditions, crop growth, and field visibility change throughout the year. We include these tasks to evaluate whether the learned embeddings are useful for broader agricultural monitoring. Following the experiment you suggested in Q2 to check whether benefit is coming from intra-annual embedding or a slightly better model; we report that the single December student embedding performs slightly below the annual teacher on cropland F1 Score (by 0.0155 in Sudan and 0.0157 in Ukraine).

Responses to Questions

[Q1] Have you done an analysis about spatial autocorrelation or have you included a spatial buffer around your 15% validation dataset?

We did not use a spatial buffer in the random 85/15 distillation split. To assess potential spatial autocorrelation, we analyzed validation performance by distance to the nearest training sample. December cosine similarity is 0.9789 within 1 km and 0.9756 at 20 – 100 km. Downstream evaluation points are separate from the distillation samples, with median distances to the nearest training point of 18 km in Sudan and 21 km in Ukraine. Downstream probes also use spatial-block cross-validation to reduce spatial leakage.

[Q2] Have you done ablations that test whether your last months embedding is actually better suited for other downstream tasks than the AEF embeddings? This would clear up if the benefit comes from the intra-annual embeddings or just a slightly better model

As you suggested, we compared the single December student embedding directly against annual AEF on downstream cropland classification. As expected the student performs slightly below the annual AEF F1 score by 0.0152 in Sudan and 0.0157 in Ukraine. This proves that the benefits are not coming from a better model. Our main contribution is that the student provides updated embeddings throughout the year, before the current-year annual AEF becomes available.

Thank you for the detailed responses and the additional analyses. The spatial autocorrelation analysis and the comparison of the single December student embedding with the annual AEF address my concerns and clarify that the benefit stems from the intra-annual updates rather than from a stronger model (minor note: the reported F1 gap for Sudan differs between the responses, 0.0155 vs. 0.0152. Please make this consistent in the revision).

While I consider this a nice and interesting contribution, I will keep my score, as I find its overall significance somewhat limited for ICLR. However, I would not stand in the way of acceptance if the concerns raised by Reviewer gSpa are resolved.

Thank you for your positive assessment and constructive feedback.

We have corrected the inconsistency in the reported Sudan F1 gap. We have also addressed Reviewer gSpa's concerns in the revised manuscript. All changes are highlighted in yellow.

Reviewer_Hs2ROpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary
  • Defined a practical problem when using earth embeddings: these embeddings are typically annual and retrospective, limiting their use for within-year tasks. Generating new embeddings can also be computationally expensive and impractical at large scales. This precludes their use for temporally sensitive tasks like agricultural monitoring, where embeddings must be available and recomputed frequently within the growing season
  • Presented an uncertainty-aware framework that distills annual embeddings into temporally evolving representations using partial satellite time series observations, showing that the resulting embeddings support in-season agricultural monitoring tasks including cropland classification and field boundary segmentation. Student embeddings achieve high cosine similarity to the annual teacher embedding even with one month of data, and downstream performance matches or exceeds the annual teacher in most tasks
  • Evaluated the approach on agricultural monitoring datasets: pixel-level cropland classification, cropland change detection, and agricultural field boundary segmentation, showing that annual representations can be recovered early in the year and used for downstream agricultural tasks, while predicted uncertainty provides a useful measure of representation reliability.
Critical strengths
  • The paper is well-written and easy to follow
  • The paper defined an existing practical problem when deploying earth embeddings: these embeddings are typically annual and retrospective, limiting their use for within-year tasks
  • The paper designed a framework with a frame encoder, a causal transformer, in-season embedding and uncertainty heads, and an auxiliary seasonal dynamics head, and provided detailed rationale for each design
  • The whole storyline is novel but natural, clearly drawing a picture from problem formulation to framework design to experiment results
Critical weaknesses
  • The proposed framework contains four different parts: a frame encoder, a causal transformer, in-season embedding and uncertainty heads, and an auxiliary seasonal dynamics head. Would make more sense to see an ablation study to show that each of them is necessary
  • AEF embeddings provide on-demand monthly (or denser) embeddings, and TESSERA monthly embeddings can be derived directly from the pretrained checkpoints; it would be good to see the comparison of the learned monthly embeddings with the actual ground truth
  • I wonder if there is a possibility that embeddings themselves don't have large gaps within the same location, so the input satellite image time series affects the downstream task more? (Figure 4)
  • Figure 5: What does Teacher AEF refer to? Previous year or current year?
  • Table 1: Why does TESSERA (previous year) perform even better than annual TESSERA on Sudan?
Questions for authors
  1. "On fields that stopped being cultivated, the previous-year AEF is correct for only 29% in Sudan and 15% in Ukraine, whereas the student’s F1 score rises through the season to 39% and 60% by December, above the annual AEF oracle (29% and 55%; Fig. 5)." -> typo? 15% and 55%?
  2. "On newly cultivated fields, the student improves in Ukraine (34% to 50%) but not in Sudan, and on unchanged fields all embeddings reach 85–97%." -> Why does this happen? Do you have any assumptions?
3 comments
AuthorsOpenReview ↗

Response (1/2)

Thank you for the constructive feedback. We address identified weaknesses and questions below:

Responses to Critical Weaknesses

[W1] The proposed framework contains four different parts: a frame encoder, a causal transformer, in-season embedding and uncertainty heads, and an auxiliary seasonal dynamics head. Would make more sense to see an ablation study to show that each of them is necessary.

We conducted ablations of the main components and proved their contribution to the method. Our model uses these configurations: patch 15×15, attention pooling, previous year conditioning, causal attention. We report cosine similarity ↑ between student and teacher (AEF) embeddings on the validation set, at the Jan, Jun and Dec cutoffs.

AblationVariantJanJunDec
Ours0.95150.96110.9748
With prior
Patch size1×10.95130.95970.9664
7×70.95330.96270.9704
31×310.95300.96090.9659
PoolingCenter0.95390.96300.9709
Center + mean0.95250.96180.9692
Prev. yearOff0.76220.89060.9090
60% available0.94540.95560.9621
Without prior
AttentionNone (single month)0.77030.75090.7599
Bidirectional (oracle)0.92290.92260.9231
Dynamics+ L_dyn0.76060.89120.9107

Previous-year conditioning and causal temporal attention give the largest gains. Patch size and pooling choices have a small effect early in the season (within 0.003 in Jan), and our configuration performs best by December (0.9748 vs. 0.9709 for the next best).

[W2] AEF embeddings provide on-demand monthly (or denser) embeddings, and TESSERA monthly embeddings can be derived directly from the pretrained checkpoints; it would be good to see the comparison of the learned monthly embeddings with the actual ground truth

TLDR: We cannot compare with monthly AEF since it is currently available only in private preview. Compared to TESSERA v1, our student outperforms early in the season and is comparable in the later months while being 50x smaller.

While AEF allows users to request temporally fine-grained embeddings, acquiring custom embeddings for each subsequent month is slow, redundant, and only available in private preview.

We ran the open TESSERA encoder (which is 50 times larger and 4.8 times slower than our student) on the observations (Sentinel-1 and Sentinel-2) available at each month, with its original preprocessing. We evaluated embedding similarity and downstream cropland classification at every monthly cutoff, comparing direct TESSERA inference with our TESSERA-distilled student.

Sudan:

CutoffTESSERA cosine ↑Student cosine ↑TESSERA F1 ↑Student F1 ↑
Jan0.64490.85500.73820.7700
Feb0.67440.86070.73140.7678
Mar0.71160.86270.75030.7694
Apr0.74830.86390.75760.7722
May0.77700.86420.76270.7711
Jun0.80670.86660.76310.7778
Jul0.82710.86920.79330.7878
Aug0.86790.87520.81160.8050
Sep0.90840.88610.82390.8290
Oct0.94140.89360.82270.8356
Nov0.96340.89870.83600.8383
Dec0.96850.89880.84110.8305

Ukraine:

CutoffTESSERA cosine ↑Student cosine ↑TESSERA F1 ↑Student F1 ↑
Sep′0.89080.84360.86020.8561
Oct′0.92670.84510.86370.8587
Nov′0.93810.84350.86630.8585
Dec′0.94170.84700.87050.8580
Jan0.28010.83020.80460.8647
Feb0.33280.83980.82000.8665
Mar0.39510.84330.83700.8652
Apr0.45780.84770.85170.8680
May0.55700.85360.87820.8769
Jun0.68530.86000.89610.8820
Jul0.77800.86970.92190.9019
Aug0.85630.87610.92700.9062
Sep0.89220.87960.93240.9143
Oct0.92800.88210.93040.9147
Nov0.93820.88050.93370.9150
Dec0.94150.88280.93200.9106

Both cosine columns measure similarity to the same annual TESSERA embeddings for the corresponding calendar year. Sep′ – Dec′ refer to the previous autumn. F1 uses the embedding at each month with the same probe protocol. Bold indicates the higher value. The student is closer to the annual teacher early in the season and achieves higher early-season F1. Direct TESSERA inference performs better by December, although the Sudan F1 difference is not statistically significant.

AuthorsOpenReview ↗

Response (2/2)

[W3] I wonder if there is a possibility that embeddings themselves don't have large gaps within the same location, so the input satellite image time series affects the downstream task more? (Figure 4)

Our goal is to recover the annual teacher representation from partial-year observations, by updating the previous year representation with in-season satellite observations. When the input satellite image time series captures a change on ground, our student embeddings diverge from the previous year teacher embeddings as seen in Figure 4. Whether the teacher itself captures the change in task-relevant information is beyond the scope of this work. One can try to fill the gap in AEF embeddings by naively combining them with raw satellite image time series. We test this against our student and our student achieves higher December F1 in Sudan (0.8479 vs. 0.8320) and Ukraine (0.9213 vs. 0.9064). This shows that our student fills the AEF gaps better.

[W4] Figure 5: What does Teacher AEF refer to? Previous year or current year?

Teacher AEF in Figure 5 means the (current year - previous year) annual AEF as the task is inter-year change detection. We will clarify this in the caption.

[W5] Table 1: Why does TESSERA (previous year) perform even better than annual TESSERA on Sudan?

These values were evaluated on different subsets, and we will fix the comparison. Precomputed annual TESSERA v1 covers only 504 of 1,796 Sudan point-seasons (https://geotessera.org/coverage). On the 504 shared samples, annual TESSERA performs better than previous-year TESSERA (0.7703 vs. 0.6849). The student achieves 0.7429, closer to the annual teacher, with a difference of -0.0275 [-0.0674, 0.0102]. We will report matched-subset results in Table 1.

Responses to Questions

[Q1] "On fields that stopped being cultivated, the previous-year AEF is correct for only 29% in Sudan and 15% in Ukraine, whereas the student’s F1 score rises through the season to 39% and 60% by December, above the annual AEF oracle (29% and 55%; Fig. 5)." -> typo? 15% and 55%?

Thanks for pointing it out, it was a typo, the actual number is 55%, we will fix it.

[Q2] "On newly cultivated fields, the student improves in Ukraine (34% to 50%) but not in Sudan, and on unchanged fields all embeddings reach 85–97%." -> Why does this happen? Do you have any assumptions?

We see that annual AEF captures a weaker cultivation-change signal in Sudan (this is related to your comment in W3). For newly cultivated fields, the distance between the previous year and current year embeddings (1 − cosine) is 0.099 in Sudan versus 0.185 in Ukraine.

AuthorsOpenReview ↗

Follow-up on our responses

Dear Reviewer Hs2R,

Thank you again for your thoughtful feedback. As the discussion period is coming to an end, we wanted to kindly follow up and ask whether our responses and additional experiments have addressed your concerns and questions.

We would greatly appreciate any further feedback.

Comments on the paper

1 comment

Dear Reviewers,

We would like to sincerely thank you for valuable feedback and suggestions. These have helped us identify where the presentation and experimental validation could be strengthened.

Our main contribution is to provide timely monthly embeddings that combine the previous year’s embedding with the current year’s monthly observations to approximate the year-end embedding. This enables sub-annual tasks such as in-season crop monitoring and crop change detection. Our approach is also cheaper than generating monthly embeddings directly with GeoFMs (TESSERA v1, OlmoEarth-Nano), while achieving comparable or better performance.

We thank both reviewers Hs2R and gSpa for asking for a comparison of our method to similar sized temporal encoders. This led us to add experiments showing that our method is comparable (or better early in the calendar-year) to TESSERA v1 and OlmoEarth-Nano while being 4.8 and 1.3 times faster in throughput (km2/skm^2/s) respectively.

We are glad that Reviewer bDAP pointed out the impact of spatial autocorrelation. We show that validation performance does not change by distance to the nearest training sample.

We thank Reviewer gSpa for suggesting to compare different aggregation schemes of our monthly embeddings to match feature dimensions to baseline. We show that our student outperforms the baseline at matched dimensionality. Their comment regarding W3 also helped us clarify that our methods can be adapted to different temporal resolution/cadence.

We responded to all the reviewer concerns by adding new experiments and clarifications. The reviews greatly added value to our work.

Authors

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.