Enhanced mapping of a smallholder crop farming landscape through image fusion and model stacking

Abstract

Globally, Smallholder farming systems (SFS) are recognized as one of the most important pillars of rural economic development and poverty alleviation because of their contribution to food security. However, support for this agricultural sector is hampered by lack of reliable information on the distributions and acreage of smallholder fields. This information is essential in not only monitoring food security and informing markets but also in guiding the determination of levels of support required from government by individual farmers. There is urgent need for robust techniques that can be used to cost-effectively and time-efficiently map smallholder crop fields especially in Sub-Saharan Africa and Asia. This study attempts to do this by using an approach in which optical and Synthetic Aperture Radar (SAR) data are systematically combined and classified using Extreme Gradient Boosting (Xgboost). We also investigated model stacking as another technique to improve classification accuracy. We combined Xgboost with Random Forest (RF), Support Vector Machine (SVM), Artificial Neural Networks (ANN), and Naïve Bayes (NB). The combined use of multi-temporal Sentinel-2 bands, spectral indices, and Sentinel-1 produced better results than exclusive use of optical data (α = 0.95, p = 0.0005). Furthermore, stacking of classification algorithms based on model comparisons achieved higher accuracy than stacking the algorithms indiscriminately (α = 0.95, p = 0.0100). Through systematic fusion of SAR and optical data and hyper-parameter tuning of Xgboost, we achieved a maximum classification accuracy of 97.71%, while achieving a maximum accuracy of 96.06% through model stacking. This highlights the importance of multi-sensor data fusion and multi-classifier systems when mapping fragmented agricultural landscapes.