In a significant boost for agricultural monitoring, a recent study has demonstrated the immense potential of machine learning in refining cropland mapping in Nigeria. The research ingeniously combined a global cropland dataset with a meticulously hand-labelled dataset, leveraging time series data from remote sensing sources like Sentinel-1 and 2, ERA5 climate data, and DEM data. Binary labels were employed to indicate the presence or absence of cropland.
Machine Learning’s Role in Agricultural Monitoring
The study meticulously employed 1827 manually labelled pixels across Nigeria, segregating them into training, validation, and test sets. Multiple machine learning models, including Long Short-Term Memory (LSTM) neural network classifiers and a Random Forest classifier, underwent evaluation against three established global land cover maps at a 10m resolution. This evaluation aimed to gauge the efficacy of these models in detecting and mapping cropland.
Striking the Right Balance
Researchers delved into the implications of incorporating or excluding the Geowiki cropland dataset to understand the delicate balance between data quantity and quality. This investigation played a pivotal role in comprehending the potential enhancements achievable by integrating diverse datasets and the influence of data quality on training machine learning models.
Findings and Their Implications
Results from the study unveiled that the existing WorldCover map demonstrated the highest F1-score and accuracy on the test set. Closely following was a single-headed LSTM model trained with the hand-labelled data and Geowiki data points specific to Nigeria. These outcomes not only accentuate the potential of machine learning in agricultural monitoring but also emphasize the critical role of selecting appropriate datasets for model training.
