PotSim: A Large-Scale Simulated Dataset for Benchmarking AI Techniques on Potato Crop

This repository contains the official implementation associated with this paper. The corresponding dataset is publicly available here.

Description:

PotSim is a large-scale simulated agricultural dataset specifically designed for AI-driven research on potato cultivation. This dataset is grounded in real-world crop management scenarios and extrapolated to approximately 4.9 million hypothetical crop management scenarios. It encompasses diverse factors including varying planting dates, fertilizer application rates and timings, irrigation strategies, and 24 years of weather data. The resulting dataset comprises over 675 million daily simulation records, offering an extensive and realistic framework for agricultural AI research.

---

Features:

The repository contains three main files example.ipynb, plots.ipynb, and run.py. To reproduce the train/test results presented in the paper, we provide run.py, which can be executed over a command line interface or terminal. To follow a step by step procedure and work with our dataset, we provide example.ipynb, a jupyter notebook template, which act as a starting point for further exploration. To make it easier to visualize and plot the results, we have provided plots.ipnb, a jupyter notebook template, which contains few example plots and can be edited according the requirements.

---

Directory Structure:

Directory Name	Description
data	Contains all datasets required for experiments and analyses.
data/potsim_yearly	Default location for yearly dataset files utilized in the study.
models	Houses all model architecture definitions and related scripts.
outputs	Default directory for saving model checkpoints, logs, and results generated during training.
saves	Stores pre-trained model states and checkpoints from experiments referenced in the paper.
testing	Includes scripts and functions for evaluating model performance and generating metrics.
training	Contains training routines, configuration files, and code for model optimization.
utils	Utility functions for data preprocessing, splitting, and model configuration management.
utils/potsimloader	Specialized utilities for efficient data loading and processing workflows.

---

Requirements:

To install the requirements:

conda env create -f environment.yml
conda activate potsim_env

Depending on the version of CUDA on your system, install PyTorch v2.5.1 from official PyTorch source at https://pytorch.org

# Example for cuda-version 12.4
pip install torch==2.5.1 torchvision==0.20.1 torchaudio==2.5.1 --index-url https://download.pytorch.org/whl/cu124

To allow on gpu metrics and display the model parameters clearly

pip install torchmetrics==1.7.1 torchinfo==1.8

If your system is not set up with conda package manager, then please visit https://www.anaconda.com/download to install Miniconda accoding to your operating system and then continue by installing the requirements from above.

---

Usage: run.py

The script supports two main commands: train and test.

• Make sure your datasets are in the .parquet format and accessible by the script at data folder.
• For more details on available target variables and models, check the code or add a --help flag:

python run.py --help
python run.py train --help
python run.py test --help

1. Train a Model

python run.py train -tgt -m  [options]

Arguments:

Argument	Type	Required	Default	Description
-tgt, --target	str	Yes		Target variable to predict. Choices: see below
-m, --model	str	Yes		Model type to use. Choices: see below
-tdata, --train_dataset	str	No	train_split	Training dataset split
-vdata, --val_dataset	str	No	val_split	Validation dataset split
-bs, --batch_size	int	No	256	Batch size
-lr, --learning_rate	float	No	0.005	Learning rate
-ep, --epochs	int	No	100	Maximum number of epochs
-sl, --seq_len	int	No	15	Sequence length (for sequence models)
-d, --device	str	No	None	Device: cpu or cuda

Example:

python run.py train -tgt="NTotL1" -m="lstm" -tdata="train_split" -vdata="val_split" -bs=256 -lr=0.001 -ep=10 -sl=15 -d="cuda"

---

2. Test a Model

python run.py test -tgt  -m  -data  [options]

Arguments:

Argument	Type	Required	Default	Description
-tgt, --target	str	Yes		Target variable to predict. Choices: see below
-m, --model	str	Yes		Model type to use. Choices: see below
-data, --dataset	str	Yes		Dataset to run test on
-mdir, --model_dir	str	No	saves	Directory where trained models are saved (outputs or saves)
-bs, --batch_size	int	No	256	Batch size
-sl, --seq_len	int	No	15	Sequence length (for sequence models)
-d, --device	str	No	None	Device: cpu or cuda

Example:

python run.py test -tgt="NTotL1" -m="lstm" -data="test_split" -mdir="saves" -bs=256 -sl=15 -d="cuda"

---

Results:

R2 Metrics for Different Models

Target	CNN1D	Transformer	LSTM	LinearRegression	MLP	TCN
NLeach	0.432	-0.02	0.343	0.002	0.014	0.265
NPlantUp	0.803	0.733	0.794	0.322	0.753	0.791
NTotL1	0.843	0.764	0.831	0.481	0.779	0.823
NTotL2	0.861	0.799	0.849	0.489	0.792	0.843
SWatL1	0.973	0.949	0.972	0.620	0.841	0.950
SWatL2	0.944	0.783	0.928	0.700	0.816	0.914