Enhancing corn future price prediction through multi-modal data fusion: a comparative analysis of LSTM, sLSTM, mLSTM, and xLSTM models

Title Multi-source Deep Learning Models for Agricultural Futures Price Prediction Description The project investigates the forecasting performance of several LSTM-based deep learning models for agricultural futures prices using multi-source data, including market data and remote sensing indicators. The repository provides the scripts used for data processing, model training, and model comparison. Dataset Information The repository includes two datasets: dataset_English.csvThe base dataset containing agricultural futures price data and related variables. dataset+remote_sensing_English.csvAn extended dataset that incorporates remote sensing indicators (e.g., NDVI and EVI) extracted from satellite imagery. Remote sensing indicators were generated using Google Earth Engine (GEE).The GEE workflow is demonstrated using Iowa (USA) in January 2021 as an example. Due to platform limitations, Google Earth Engine does not support exporting the full computational environment or interactive scripts as standalone files.Therefore, the reproducible workflow is shared through a permanent GEE script link provided in: 'google_earth_engine.txt' Code Information The repository contains the following scripts: correlation index.pyScript used to compute correlation indicators among variables. time_prediction.pyMain script used for model training and time-series prediction. four_models_comparison.pyScript for comparing the prediction performance of multiple deep learning models. Requirements The code was tested under the following environment: Python == 3.9torch == 2.10.0numpy == 2.2.6pandas == 2.3.3matplotlib == 3.10.8scikit-learn == 1.7.2tqdm == 4.67.1 The full Python environment used in this study can be reproducedusing the provided 'requirements_freeze.txt' file generated from 'pip freeze'. Usage Instructions To reproduce the experimental results reported in the study, run the following scripts: Experiment 1 and Experiment 2 (model comparison experiments):python four_models_comparison.py Experiment 3 (prediction experiment):python time_prediction.py Methodology The study applies deep learning models to agricultural futures price prediction using multi-source data.The workflow includes the following steps: Data collection and preprocessing Remote sensing index extraction using Google Earth Engine Construction of time-series datasets Training LSTM-based prediction models Model comparison and evaluation