Mapping Maize Cropland and Land Cover in Semi-Arid Region in Northern Nigeria Using Machine Learning and Google Earth Engine

Abstract

The monitoring of crop quantity and quality is vital for global food security. National food security has recently been at the forefront of local and regional research, and has become a vital priority for most developing countries. Therefore, ensuring reliable classification of cropland and other land cover is crucial for sustainable agricultural development and ensuring national food security. A good understanding of the Nigerian agricultural sector is essential to making better decisions and managing operations more efficiently. Scientists, practitioners, and policymakers must exchange reliable information to develop and support agricultural programs and policies. It is essential to develop and implement novel methods for mapping maize cropland and other land cover types. Thus, Seasonal Crop Inventory (SCI) is a valuable tool for farmers, researchers, and policymakers, as it provides critical information on crop production. It informs decisions related to land management, food security, and agricultural policy. In this study, Sentinel-1 and Sentinel-2 images have been combined to map maize cropland and other land covers in northern Nigeria during the 2016–2019 growing season. We employed a technologically advanced space-based remote sensing technique. As a pioneer study that obtained detailed information on northern Nigeria’s cropland, the research utilized platforms such as Google Earth Engine (GEE), a cloud-computing engine using various classification techniques that include Random Forest (RF), Support Vector Machine (SVM), and Classification Regression Trees (CART) algorithms to produce a pixel-based Seasonal Crop Inventory of the study area. The outcome demonstrated a reliable GEE-based mapping of the region’s cropland with satisfactory classification accuracy. It revealed the overall accuracy values and the Kappa coefficients to be above 97% during the different time nodes under study. It also indicated a 98% and 93% producer and user accuracy for the cropland. The research further revealed that the Random Forest performed the best among the three machine-learning models tested in this study for mapping the maize cropland and other land cover classes. Therefore, the study’s findings and the derived crop mapping would greatly help provide valuable information that helps farmers, policymakers, and other stakeholders make more informed decisions about agricultural production, land use planning, and resource management.