Mapping sugarcane plantations in Northeast Thailand using multi-temporal data from multi-sensors and machine-learning algorithms

The effectiveness of machine learning algorithms and the limited reference data introduce uncertainness for sugarcane classification. To address these problems, our study classified sugarcane plantations at the field scale using multi-temporal and multi-sensor data together with a large number of ground truth datasets (>13,000 points) and compared the efficacy of ensemble and kernel classifier methods over 3 years (2021, 2022, and 2023) across Northeast Thailand. In the first step, land cover was generated from a random forest classifier, demonstrating excellent results for all years with an OA higher than 95%. In the second step, the discretization of sugarcane from non-sugarcane classes in the agricultural category was conducted using four efficient machine learning algorithms (decision tree (DT), random forest (RF), support vector machine (SVM), and one-class SVM). The RF classifier gave the optimal results with over 90% accuracy. Our results aligned with provincial statistics from the Office of the Cane and Sugar Board, thereby highlighting the efficacy and reliability of the RF method in mapping sugarcane in small fields and cloudy regions. A temporal evolution analysis of sugarcane cultivation spanning the preceding 3 years revealed a significant increase in the productive area. Our findings provide crucial information for sustainable management practices.