UAV-based mapping of banana land area for village-level decision-support in Rwanda

• Published in: Remote sensing (Basel, Switzerland) Switzerland), 2021-12, Vol.13 (24), p.4985
• Main Authors: Kilwenge, Regina, Adewopo, Julius, Sun, Zhanli, Schut, Marc
• Format: Article
• Language: English
• Subjects: Accuracy; Agricultural land; Agricultural practices; Agricultural production; Agriculture; Agronomic crops; banana; BXW; Cash crops; Classification; Classifiers; Climate adaptation; Climate change; Corn; Coverage; Crop production; Crops; Cultivated lands; Cultivation; Decision support systems; Decision trees; Farmers; Farms; Food security; Fruit cultivation; Fruits; Land area; Land cover; land cover mapping; Learning algorithms; Machine learning; Pests; precision agriculture; Regression analysis; Remote sensing; Resource allocation; Rwanda; Spatial distribution; Support vector machines; UAV; Unmanned aerial vehicles; Vegetation; Wilt
• ISSN: 2072-4292; 2072-4292
• DOI: 10.3390/rs13244985

Crop monitoring is crucial to understand crop production changes, agronomic practice decision-support, pests/diseases mitigation, and developing climate change adaptation strategies. Banana, an important staple food and cash crop in East Africa, is threatened by Banana Xanthomonas Wilt (BXW) disease. Yet, there is no up-to-date information about the spatial distribution and extent of banana lands, especially in Rwanda, where banana plays a key role in food security and livelihood. Therefore, delineation of banana-cultivated lands is important to prioritize resource allocation for optimal productivity. We mapped the spatial extent of smallholder banana farmlands by acquiring and processing high-resolution (25 cm/px) multispectral unmanned aerial vehicles (UAV) imageries, across four villages in Rwanda. Georeferenced ground-truth data on different land cover classes were combined with reflectance data and vegetation indices (NDVI, GNDVI, and EVI2) and compared using pixel-based supervised multi-classifiers (support vector models-SVM, classification and regression trees-CART, and random forest-RF), based on varying ground-truth data richness. Results show that RF consistently outperformed other classifiers regardless of data richness, with overall accuracy above 95%, producer's/user's accuracies above 92%, and kappa coefficient above 0.94. Estimated banana farmland areal coverage provides concrete baseline for extension-delivery efforts in terms of targeting banana farmers relative to their scale of production, and highlights opportunity to combine UAV-derived data with machine-learning methods for rapid landcover classification.