Lite Learning: Efficient Crop Classification in Tanzania Using Traditional Machine Learning & Crowd Sourcing

Abstract

This study introduces a novel methodology for crop type classification in Tanzania by integrating crowdsourced data with time-series features extracted from Sentinel-2 satellite imagery. Leveraging the YouthMappers network, we collected ground validation data on various crops, including challenging types such as cassava, millet, sunflower, sorghum, and cotton across a range of agricultural areas. Traditional machine learning algorithms, augmented with carefully engineered time-series features, were employed to map the different crop classes. Our approach achieved high classification accuracy, evidenced by a Cohen's Kappa score of 0.80 and an F1-micro score of 0.82. The model often match or outperform broadly used land cover models which simply classify 'agriculture' without specifying crop types. By interpreting feature importance using SHAP values, we identified key time-series features driving the model's performance, enhancing both interpretability and reliability. Our findings demonstrate that traditional machine learning techniques, combined with computationally efficient feature extraction methods, offer a practical and effective “lite learning” approach for mapping crop types in data-scarce environments. This methodology facilitates accurate crop type classification using a low-cost, resource-limited approach that contributes valuable insights for sustainable agricultural practices and informed policy-making, ultimately impacting food security and land management in resource-limited contexts, such as sub-Saharan Africa.

Keywords: Agriculture, Feature Extraction, Classification, Time-Series, Crop Type