Simultaneous multi-crop land suitability prediction from remote sensing data using semi-supervised learning.

Area of Science:

• Agricultural Science
• Machine Learning
• Geospatial Analysis

Background:

• Current Canadian land suitability models rely on single-crop data and expert judgment, limiting their scope and accuracy.
• There is a need for advanced, data-driven approaches to assess suitability for diverse crops simultaneously.

Purpose of the Study:

• To develop and validate a novel data-driven, multi-crop land suitability model for Canada.
• To improve the accuracy and scope of agricultural land suitability predictions using machine learning.

Main Methods:

• A multi-layer perceptron (MLP) model was employed, utilizing downscaled farm-level crop yield data (2013-2020).
• Soil, climate, and landscape variables from Google Earth Engine were integrated.
• A semi-supervised learning approach with a crop indicator function accommodated varying data resolutions and unlabeled data, capturing inter-crop dependencies.

Main Results:

• The multi-crop model achieved up to a 2.82-fold reduction in mean absolute error compared to single-crop models.
• Barley, oats, and mixed grains demonstrated higher tolerance to environmental variations, suitable for wider Canadian regions.
• Non-grain crops showed greater sensitivity to environmental factors.

Conclusions:

• The developed model offers a significant improvement over traditional methods for predicting land suitability for multiple crops.
• Findings indicate increasing agricultural potential in Northern Canada due to climate change, supporting land-use assessments.
• The model can aid in evaluating northern land suitability for cultivation and inform cost-benefit analyses.