This repository for an experimental projection model which aims to project broadacre harvest crop yield based on limited reference data and without ground-truth sampling.

Dryland winter cropping refers to the cultivation of crops such as wheat, barley, canola, lupins, and pulses which are not irrigated and are dependent on rainfall from late-Autumn through Winter. This model will test the relationship between vegetation indices and crop yield from 1989 to 2020. There has been a growing field of academic research on correlating these variables at the paddock, local and small regional level, but not allot that attempts to predict for large volumes of land cover, such as the larger regional or state level.

With open-source programming and cloud computing technologies becoming more accessible, I was motivated to attempt a general remote sensing approach to the question: how well can we predict the next harvest from 6 + months out? Because of both systemic long-term climate change and increasing frequency of severe weather events (bushfires, flooding, drought), being able to predict crop yield earlier in the winter season is becoming more commercially and environmentally important.

The model tests a the relationship between both EVI (Enhanced Vegetation Index) and NDVI (Normalized Difference Vegetation Index) to test for the best vegetation model fit, then tests other spectral variables, NBR (Burn Ratio) and NDSI (Snow Index) into the model to see how closely a multiple regression model predicts to current published estimates for the AOI 2021 harvest.

• pandas
• numpy
• datetime
• re
• jupyter
• dload

• scipy
• statsmodels.api
• sklearn

• geopandas
• rasterio
• rioxarray
• xarray
• shapely

• matplotlib
• folium
• seaborn

• ee
• eemont

The project includes a function file which includes custom functions which are imported into the working calculation sheets

• Landsat 5 imagery sourced through the Google Earth Engine
• Landsat 7 imagery sourced through the Google Earth Engine
• Longpaddock SILO weather data (gridded raster weather data to 1ha squared resolution - Link
• ABARES (Australian Bureau of Agricultural and Resource Economics and Sciences) Classification and Land Use Raster (2018) - Link
• ABARES Australian Crop Report - Link

1. py01_helper_functions.py (Functions)

   • This file sets out all the functions required to run the model.

2. py02_model_data_setup.py

   • This file sets up the non-cloud data needed for the model such as feature engineering of the ABARES crop report and import and formatting of the land classification geodataframe

3. py03_get_silo.py

   • This file pulls the weather data from the SILO AWS server. The weather data is not sued in the primary analysis but had been included for future random forest model testing.

4. py04_earth_engine.py

   • This file

5. py05_regression_modelling.py

   • This file

6. py06_random_forrest.py

   • File is a test model This file tests a workflow for formatting the data into a random forest panel table and runs a simple scikit regression model across all the features

7. py07_mapping.py

   • This file is used for mapping data and printing images from the analysis

To set up the environment for this model, with Git and Miniconda/Anaconda installed, run the following in the command terminal (this model is tested using Bash)

git clone https://github.com/mike-petrut/crop-yield-forecaasting
cd crop-yield-forecasting
conda env create --file environment.yml
conda activate crop-yield-forecasting

Then install the required packages into the environment using

pip install -r requirements.txt

Once the model is set up, the python files can be run in order using the following code in Bash

for f in *.py; do python "$f"; done

The model will ask for an authentication key and open your default browser, prompting you to log onto your Gmail account which has an associated Google Earth Engine key. The authentication code provided in Google can then be copy-paste into Bash and the model will continue to run.

To run the blog post and generate a HTML file, run the blog_post_final_210705.ipynb from start to finish

An example of how to run this model will be to define a link to the geodataframe that you would like to be the area of interest. The file used in this document is all the croplands for South Australia, but this can be reproduced using another region of Australia. Note that the actual data for yield used is for South Australia, this can be changed to other states or territories by changing the test on line 15 of 'py02_model_data_setup'. If the user wants to test a geography outside Australia they will need to add their data source of historical values into the model in this same notebook.

Land use data for other states and territories are available from this link. To run the model as is, set up for South Australia, simply run the py00_run_full_model.py file in the repository. ​

• Source more local and regional time-series data from government and industry groups to test the model hypothesis across multiple regions incorporating soil data, elevation, and other geographic variables.
• Experiment with random forest models to further evaluate the impact each month throughout the year has on the final harvest yield.
• Explore more spectral band combinations using eemont that can be used as input variables.