Conference Agenda

Multi-modal Satellite Image Time Series (SITS) analysis faces significant computational challenges for live land monitoring applications. While Transformer architectures excel at capturing temporal dependencies and fusing multi-modal data, their quadratic computational complexity and the need to reprocess entire sequences for each new acquisition limit their deployment for regular, large-area monitoring. This paper studies various dual-form attention mechanisms for efficient multi-modal SITS analysis, that enable parallel training while supporting recurrent inference for incremental processing.

We compare linear attention and retention mechanisms within a multi-modal spectro-temporal encoder. To address SITS-specific challenges of temporal irregularity and unalignment, we develop temporal adaptations of dual-form mechanisms that compute token distances based on actual acquisition dates rather than sequence indices.

Our approach is evaluated on two tasks using Sentinel-1 and Sentinel-2 data: multi-modal SITS forecasting as a proxy task, and real-world solar panel construction monitoring. Experimental results demonstrate that dual-form mechanisms achieve performance comparable to standard Transformers while enabling efficient recurrent inference. The multi-modal framework consistently outperforms mono-modal approaches across both tasks, demonstrating the effectiveness of dual mechanisms for sensor fusion. The results presented in this work open new opportunities for operational land monitoring systems requiring regular updates over large geographic areas.

Sensor harmonization is required to produce consistent Landsat imagery for long-term change detection. This study investigated the effect of seasonality and aerosol optical thickness (AOT) on linear harmonization functions, which are frequently used to create consistent Landsat 7 ETM+ and Landsat 8 OLI time series data. We found that training harmonization functions with pixels that have low or average AOT can greatly reduce the difference between near-coincidental Landsat 7 and Landsat 8 observations, and that seasonally trained harmonization models outperform models trained on year-round data. We assessed the effect of ETM+/OLI sensor harmonization on forest type classification with a Random Forest model, and found that seasonally harmonized imagery provided more consistent classification maps than the alternatives. This study illustrates important details related to the creation of harmonized datasets and is a significant step toward creating more consistent Landsat 7 and Landsat 8 data for long-term change detection analysis.

Urban growth in Ecuador's inter-Andean valleys has accelerated the territory's transformation, driven by the expansion of road infrastructure and the occupation of environmentally fragile areas. In this context, the Ruta Viva highway has reconfigured urbanisation patterns in the parishes of Cumbaya and Tumbaco, advancing the urban frontier into agricultural areas and moderate slopes. The objective of this study is to evaluate the dynamics of urban expansion in the parishes of Cumbaya and Tumbaco during the period 2002-2032, using a multitemporal and predictive approach to project future urbanisation scenarios and generate inputs for sustainable territorial planning and land management. The methodology integrated multitemporal analysis of land use and land cover data from MapBiomas (2002-2022), predictive modelling using CA-Markov-MOLUSCE, and urban expansion analysis. The results show a 3% increase in urban coverage during the 2002-2022 period and a projected 12% growth by 2032, concentrated south of the Ruta Viva corridor and within the agricultural mosaic. Simulations show that slopes below 25° are more susceptible to urbanisation, while vegetation cover loss reaches 30% on the slopes of Ilalo Hill. This study provides a robust, replicable tool for anticipating urbanisation scenarios in Andean environments, guiding land management and environmental conservation strategies in regions of high urban pressure.

The Area-Based Approach (ABA) for modelling forest biomass with ALS data assumes perfect co-registration, but operational inventories often have spatiotemporal misalignments. This study isolates and quantifies the independent error contributions from temporal gaps and spatial co-location errors. The analysis uses a unique dataset from the Taiga Plains, Canada, featuring 163 re-measured field plots paired with repeated ALS acquisitions from the same sensor.

To assess temporal effects, we constructed scenarios with varying time-gap distributions. Symmetrical time gaps (SD 1.1 vs 2.5 years) increased RMSE by ~1 percentage point but did not add bias. In contrast, skewed distributions introduced significant systematic biases of 8.0 % (6.8 Mg ha⁻¹).

To assess spatial effects, we linked co-location uncertainty directly to plot-level neighbourhood heterogeneity. This was done by shifting the 20x20m ALS footprint over a 1m lattice and recalculating predictors. The resulting predictor variability (RMS(CV) 12.7%) was propagated through the model, implying a positional sigma of 10-15%. Monte Carlo simulations confirmed this spatial component is the dominant error source, contributing 2–4 percentage points to the ~22% baseline %RMSE.

Our findings show that while balanced temporal gaps are manageable, spatial co-location affected by the local heterogeneity is the most critical factor for robust ABA models.