Cropland Suitability Assessment Using Satellite-Based Biophysical Vegetation Properties and Machine Learning

The determination of cropland suitability is a major step for adapting to the increased food demands caused by population growth, climate change and environmental contamination. This study presents a novel cropland suitability assessment approach based on machine learning, which overcomes the limitations of the conventional GIS-based multicriteria analysis by increasing computational efficiency, accuracy and objectivity of the prediction. The suitability assessment method was developed and evaluated for soybean cultivation within two 50 x 50 km subsets located in the continental biogeoregion of Croatia, in the four-year period during 2017-2020. Two biophysical vegetation properties, leaf area index (LAI) and a fraction of absorbed photosynthetically active radiation (FAPAR), were utilized to train and test machine learning models. The data derived from a medium-resolution satellite mission PROBA-V were prime indicators of cropland suitability, having a high correlation to crop health, yield and biomass in previous studies. A variety of climate, soil, topography and vegetation covariates were used to establish a relationship with the training samples, with a total of 119 covariates being utilized per yearly suitability assessment. Random forest (RF) produced a superior prediction accuracy compared to support vector machine (SVM), having the mean overall accuracy of 76.6% to 68.1% for Subset A and 80.6% to 79.5% for Subset B. The 6.1% of the highly suitable FAO suitability class for soybean cultivation was determined on the sparsely utilized Subset A, while the intensively cultivated agricultural land produced only 1.5% of the same suitability class in Subset B. The applicability of the proposed method for other crop types adjusted by their respective vegetation periods, as well as the upgrade to high-resolution Sentinel-2 images, will be a subject of future research.