Understanding Cotton Yield Drivers Using xLSTM

Official implementation of the study:

“xLSTM for Multi-Source Cotton Yield Estimation and Temporal Interpretability Across Agro-Ecological Regions in Türkiye”
Advances in Space Research, 2026

Furkan Yardımcı, Mustafa Serkan Isik, Alp Ertürk, and Esra Erten

---

📄 Overview

This repository contains the implementation of an interpretable deep learning framework for field-scale cotton yield prediction. The proposed xLSTM model leverages multi-source Earth Observation (EO) time series and environmental data to model spatiotemporal variability in yield.

⚠️ At this stage, only the dataset and model code are shared.
Model outputs, performance figures, and scientific conclusions are intentionally excluded to prevent unauthorized use prior to publication.

🔄 Model Workflow Overview

---

⚙️ Installation

conda create -n cottonxlstm python=3.9
conda activate cottonxlstm
pip install -r requirements.txt

No manual preprocessing is required. Preprocessed .npy or .csv files are included in the repository.

---

💻 Usage

Jupyter notebooks for model training and analysis are located in the notebooks/ directory:

• xlstm.ipynb: xLSTM model training and evaluation
• lstm.ipynb: Baseline LSTM
• bilstm.ipynb: Bidirectional LSTM
• informer.ipynb: Informer model

cd notebooks/
jupyter notebook

Each notebook loads preprocessed data from the dataset/ folder.

---

📦 Dataset Description

This dataset enables field-scale cotton yield estimation across Turkey by integrating dynamic and static environmental variables into a multivariate time-series structure.

🌍 Spatial and Temporal Scope

• Regions: Aegean, Mediterranean, Southeastern Anatolia — Turkey’s primary cotton zones
• Years: 2019–2023 cotton seasons
• Target: Annual cotton yield (kg/da), obtained from TUIK at commune level and assigned to field polygons via spatial overlay

Each sample corresponds to one field and is represented as a multivariate time series tensor. Features are aggregated into bi-monthly intervals (early and late parts of each month) across the phenology window (May–November), yielding fixed-length sequences for deep learning models.

🗺️ Regional Yield Variation

📊 Yield Distribution by Region and Year

---

🛰️ Feature Overview

Feature Type	Data Source	Temporal Resolution	Spatial Resolution
Radar backscatter (VV, VH)	Sentinel-1	12 days → daily interpolated	10 m
Enhanced Vegetation Index (EVI)	Sentinel-2	5 days → daily interpolated	20 m
Meteorological (d2m, temperature, VWC, radiation, etc.)	ERA5-Land	3-hourly → daily aggregated	~9 km
Soil properties (clay, sand, SOC, pH, etc.)	SoilGrids	Static	250 m
Yield labels	TUIK	Annual	Commune polygon

All EO features were extracted via Google Earth Engine (GEE) using annual national cotton masks.

---

⚙️ Preprocessing Pipeline

• Sentinel-1: Terrain correction using SRTM; speckle reduction via 7×7 boxcar filter
• Sentinel-1 & Sentinel-2: Interpolated to daily values, smoothed with Savitzky-Golay filter
• ERA5-Land: Hourly variables aggregated into daily statistics
• SoilGrids: Static features repeated across time to align with dynamic inputs
• All inputs structured as X ∈ ℝ^d×m, where d is the number of features and m is the number of time steps

---

🌾 Agronomic Motivation

• Bi-monthly temporal aggregation captures key growth stages (e.g. flowering, boll development)
• Feature design allows modeling of climatic, biophysical, and edaphic influences on yield
• Regional diversity (soil types, farming practices, climate) enables robust and generalizable learning
• Dataset structure supports both high-accuracy prediction and explainable model analysis

This dataset enables interpretability-focused deep learning for agricultural forecasting by aligning rich EO data with temporal field-level outcomes.

📌 Reference for data methodology:
If you use this repository, please cite:

@article{YARDIMCI2026,
  title = {xLSTM for Multi-Source Cotton Yield Estimation and Temporal Interpretability Across Agro-Ecological Regions in Türkiye},
  journal = {Advances in Space Research},
  year = {2026},
  issn = {0273-1177},
  doi = {10.1016/j.asr.2026.05.025},
  url = {https://www.sciencedirect.com/science/article/pii/S0273117726006599},
  author = {Furkan Yardımcı and Mustafa Serkan Isik and Alp Ertürk and Esra Erten}
}

---

🔗 Related Resources

• Informer (Zhou et al., 2020)
• xLSTM (Hochreiter Lab)
• LSTM Baseline
• BiLSTM Baseline

---

📬 Contact

For questions, feel free to contact:
📧 furkanyardimci1006@gmail.com