Time series feature extraction and crop type classification on spatiotemporal soil sensor network data

Understanding the dynamics of soil conditions is essential for optimizing agricultural practices, particularly for precision agriculture. This paper analyzes high-resolution soil moisture data from a spatiotemporal sensor network dataset. Leveraging the Python library tsfresh, we systematically extract time series features from overlapping windows (in time) in order to characterize the temporal evolution of soil moisture conditions. In particular, we quantify and analyze the relevance of the extracted features with respect to various crop types via statistical hypothesis testing. Based on the resulting relevancy scores, we identify subsets of informative features for training and evaluating multiple machine-learning classifiers for crop type classification. Our first results of our feature importance analysis indicate that specific time series features, particularly those associated with time-frequency decomposition and entropy measures, are linked to crop type. Moreover, our model evaluations suggest that crop type prediction is feasible using time series features derived from high-resolution soil sensor data at field scale.