Mapping sugarcane in complex landscapes by integrating multi-temporal Sentinel-2 images and machine learning algorithms

Abstract

Sugarcane is an important type of cash crop and plays a crucial role in global sugar production. Clarifying the magnitude of sugarcane planting will likely provide very evident supports for local land use management and policy-making. However, sugarcane growth environment in complex landscapes with frequent rainy weather conditions poses many challenges for its rapid mapping. This study thus tried and used 10-m Sentinel-2 images as well as crop phenology information to map sugarcane in Longzhou county of China in 2018. To minimize the influences of cloudy and rainy conditions, this study firstly fused all available images in each phenology stage to obtain cloud-free remote sensing images of three phenology stage (seedling, elongation and harvest) with the help of Google Earth Engine platform. Then, the study used the fused images to compute the normalized difference vegetation index (NDVI) of each stage. A three-band NDVI dataset along with 4000 training samples and 2000 random validation samples was finally used for sugarcane mapping. To assess the robustness of the three-band NDVI dataset with phenological characteristics for sugarcane mapping, this study employed five classifiers based on machine learning algorithms, including two support vector machine classifiers (Polynomial-SVM and RBF-SVM), a random forest classifier (RF), an artificial neural network classifier (ANN) and a decision tree classifier (CART-DT). Results showed that except for ANN classifier, Polynomial-SVM, RBF-SVM, RF and CART-DT classifiers displayed high accuracy sugarcane resultant maps with producer’s and user’s accuracies of greater than 91%. The ANN classifier tended to overestimate area of sugarcane and underestimate area of forests. Overall performances of five classifiers suggest Polynomial-SVM has the best potential to improve sugarcane mapping at the regional scale. Also, this study observed that most sugarcane (more than 75% of entire study area) tends to grow in flat regions with slope of less than 10°. This study emphasizes the importance of considering phenology in rapid sugarcane mapping, and suggests the potential of fine-resolution Sentinel-2 images and machine learning approaches in high-accuracy land use management and decision-making.