State-Space Modeling With Mamba For Interpretable Crop Yield Estimation: A Cotton Case

Accurate and reliable crop yield estimation is essential for optimizing agricultural decision-making. However, the uncertainty and variability inherent in multivariate data make this a challenging task. In this study, we explore the potential of Mamba, a recently proposed state-space model designed for long sequence modeling, for cotton yield estimation from multivariate time series data. Mamba is evaluated against established deep learning architectures, namely LSTM, BiLSTM, and the transformer-based Informer, on a cotton yield dataset composed of optical and SAR satellite imagery, meteorological variables, and static soil parameters collected from three diverse agricultural regions across Türkiye. Results demonstrate that Mamba achieves competitive predictive performance, while consistently emphasizing critical phenological periods aligned with agronomic expectations, additionally offering highly efficient inference and moderate training times, making it well-suited for national scale agricultural applications. These findings highlight Mamba's potential as a scalable, accurate, and interpretable alternative to conventional models for crop yield estimation.