Akademik Veri Yönetim Sistemi
Advances in Space Research, ss.1-18, 2026 (Scopus)
The increasing availability of complementary Earth Observation (EO) data, particularly from spaceborne optical and SAR sensors such as Sentinel-1 and Sentinel-2, has made crop yield estimation more data-driven. However, the relationship between EO features and yield can vary significantly throughout the crop growth cycle, depending on geographical, physiological, and agronomic conditions. Robust deep learning (DL) models for yield estimation must achieve not only high accuracy, but also temporally coherent relationships among features and modalities. This study addresses these challenges using extended Long Short-Term Memory (xLSTM), which extends and stabilizes traditional LSTM, a widely accepted approach to handling long time-series data in yield estimation, by improving memory retention and gradient flow over time steps. Using a multivariate time-series (MTS) dataset consisting of spaceborne EO data (Sentinel-1 SAR and Sentinel-2 optical), satellite-informed reanalysis variables (ERA5-Land), and soil data (SoilGrids), for cotton yield estimation in Türkiye, encompassing diverse environmental conditions and agricultural practices, xLSTM is compared with standard LSTM, Bidirectional LSTM (BiLSTM), and the Transformer-based Informer model. The results show that xLSTM not only achieves comparable predictive performance but also provides highly consistent temporal importance profiles across diverse sub-regions with varying agricultural practices and climatic conditions. It demonstrates strong temporal coherence and broad alignment with key agronomic stages, particularly highlighting late summer as a critical period for cotton yield. In contrast, LSTM, BiLSTM, and Informer models exhibit less consistent temporal importance patterns. This work highlights the potential of xLSTM for robust and interpretable crop yield estimation across diverse agroecological contexts. By leveraging MTS, the proposed approach supports evaluation of model robustness across heterogeneous agro-ecological settings within the study area, offering a basis for data-driven agricultural decision-making. The study contributes to the development of scalable methods for crop yield estimation based on satellite EO data, supporting regional agricultural monitoring from space. The availability of both the dataset and source code aims to support reproducibility and foster further research in yield estimation. All resources can be accessed at: https://github.com/Feanor1021/Cotton-Yield-Forecast-2025.