Reducing the labeling burden in time-series mapping using Common Ground: a semi-automated approach to tracking changes in land cover and species over time

Reliable classification of Earth Observation data depends on consistent, up-to-date reference labels. However, collecting new labelled data at each time step remains expensive and logistically difficult, especially in dynamic or remote ecological systems. As a response to this challenge, we demonstrate that a model with access to reference data solely from time step t0 can perform competitively on both t0 and a future time step t1, outperforming models trained separately on time-specific reference data (the gold standard). This finding suggests that effective temporal generalization can be achieved without requiring manual updates to reference labels beyond the initial time step t0. Drawing on concepts from change detection and semi-supervised learning (SSL), the most performant approach, “Common Ground”, uses a semi-supervised framework that leverages temporally stable regions-areas with little to no change in spectral or semantic characteristics between time steps-as a source of implicit supervision for dynamic regions. We evaluate this strategy across multiple classifiers, sensors (Landsat-8, Sentinel-2 satellite multispectral and airborne imaging spectroscopy), and ecological use cases. For invasive tree species mapping, we observed a 21-40% improvement in classification accuracy using Common Ground compared to naive temporal transfer, where models trained at a single time step are directly applied to a future time step. We also observe a 10 -16% higher accuracy for the introduced approach compared to a gold-standard approach. In contrast, when broad land cover categories were mapped across Europe, we observed a more modest 2% increase in accuracy compared to both the naive and gold-standard approaches. These results underscore the effectiveness of combining stable reference screening with SSL for scalable and label-efficient multi-temporal remote sensing classification.

💡 Research Summary

The paper addresses the persistent challenge of label degradation in time‑series remote sensing classification, where reference data collected at an initial time step (t₀) become outdated as ecosystems evolve. Collecting fresh ground truth for each subsequent acquisition (t₁, t₂, …) is costly, time‑consuming, and often infeasible, especially in remote or protected areas. To mitigate this, the authors propose a semi‑automated framework called “Common Ground” that leverages only the original t₀ labels while still achieving competitive performance on future imagery.

The core concept rests on identifying “temporally stable regions” – pixels whose spectral and semantic characteristics show little or no change between t₀ and t₁. Stability is detected automatically using change‑detection algorithms such as Iteratively Re‑weighted Multivariate Alteration Detection (IRMAD) for the invasive‑species case studies and Continuous Change Detection (CCD) for the European land‑cover case. These stable pixels are assumed to retain their original class, allowing the t₀ labels that fall within them to be reused without modification.

The workflow consists of two stages. In Stage 1, a classifier is trained on the union of (i) all t₀ reference samples and (ii) the subset of t₁ reference samples that have been confirmed stable. This model captures the shared representation of both time steps while respecting the limited, high‑confidence training set. In Stage 2, the Stage 1 model is applied to the remaining (changed) pixels in t₁; predictions with high confidence are treated as pseudo‑labels. The final model is then retrained on the combined set of (a) original t₀ labels, (b) stable t₁ labels, and (c) pseudo‑labels for changed areas. By doing so, the approach expands label coverage without any manual relabeling at t₁, effectively turning the stable region mask into an implicit supervision signal for semi‑supervised learning.

Five labeling strategies are compared across three case studies and two classifier families (Random Forest and the transformer‑based TabPFN v2). The strategies are: (1) Gold Standard – full manual relabeling at t₁; (2) Naïve Transfer – direct application of a t₀‑trained model to t₁; (3) Spectral‑Stability Sampling – pseudo‑labels drawn from a t₀ classification map within stable areas; (4) Cross‑Temporal Stable Points – use only the original t₀ labels plus the stable t₁ labels; and (5) Common Ground – the full two‑stage SSL pipeline described above.

Results reveal that for fine‑grained invasive‑tree species mapping (both Sentinel‑2 multispectral and airborne hyperspectral data), Common Ground outperforms Naïve Transfer by 21 %–40 % in overall accuracy and even surpasses the Gold Standard by 10 %–16 %. This demonstrates that, when class boundaries are narrow and temporal change is substantial, the semi‑supervised augmentation of stable regions yields a clear advantage. In contrast, for broad‑scale European land‑cover classification, the improvement is modest (≈2 % over both Naïve Transfer and Gold Standard), reflecting the lower class complexity and higher proportion of stable pixels already present in the dataset.

The study highlights several practical benefits. First, the method is lightweight and compatible with traditional machine‑learning models as well as modern tabular foundation models, avoiding the heavy computational demands of deep domain‑adaptation techniques. Second, it requires only a single set of field observations, making it attractive for monitoring programs with limited budgets or personnel. Third, by relying on change‑detection masks rather than generating new training labels from model predictions, it reduces the risk of error propagation that plagues some self‑training approaches.

Limitations are acknowledged. The effectiveness of Common Ground hinges on the quality of the change‑detection step; inaccurate masks can either discard useful training samples or mistakenly retain mislabeled ones. In ecosystems experiencing rapid, large‑scale disturbances (e.g., fire, flood), the proportion of stable pixels may become too small to sustain robust SSL, diminishing gains. Moreover, the approach does not explicitly address class‑specific phenological shifts that may alter spectral signatures without causing a change‑mask signal.

Future work suggested includes integrating multi‑temporal, multi‑sensor fusion to enrich the stable‑region signal, exploring more sophisticated uncertainty‑aware pseudo‑labeling, and extending the framework to handle severe class imbalance. The authors also propose testing the method on longer time series and operational monitoring pipelines to assess scalability and robustness over multiple successive epochs.

In summary, “Common Ground” offers a pragmatic, label‑efficient solution for temporal transfer in remote sensing classification. By exploiting spectrally stable areas as a source of implicit supervision and coupling this with a simple semi‑supervised learning loop, the framework achieves substantial accuracy gains in dynamic, fine‑scale ecological applications while keeping labeling effort to a single field campaign. This positions it as a valuable tool for conservation practitioners, land‑management agencies, and researchers seeking cost‑effective, repeatable Earth‑observation analyses.