Sugarcane Area Mapping with Zero Labels Using Hybrid Unsupervised and Supervised Approach

Mapping sugarcane areas is vital for applications like crop monitoring, yield estimation, environmental monitoring, and land use planning. Traditional supervised learning is hindered by the costly and time-intensive collection of ground truth data, whereas, unsupervised methods encounter performance challenges due to parameter initialization. In this study, we propose a hybrid approach that integrates unsupervised learning with domain knowledge of temporal sugarcane crop descriptors, creating reference data for subsequent supervised learning. The main objective of the study was to map sugarcane areas in the absence of ground reference data by combining unsupervised and supervised learning using Sentinel-1 and 2 observations. The study was carried out in two provinces in central Thailand during the sugarcane season of 2021–2022. X means clustering was applied to the raster stack of temporal NDVI and radar backscatter (in VH polarisation). Fifty polygons were then digitized from each cluster and categorized as sugarcane, cassava, field crops, and non-agriculture using the temporal descriptors for each polygon. Polygons not meeting the temporal descriptors based criteria were removed. Further, RF based supervised classification was carried out using bands of Sentinel-1 and 2, and NDVI as the features and labled polygons as the reference. The samples were divided into training (80% data) and testing (20% data). We tuned the RF classifier for a number of trees ranging from 50 to 500. The model with 350 trees performed better on testing data, with an overall accuracy of 87.75% and a Kappa of 0.873. Slight intermixing between sugarcane and cassava was observed mainly due to the planting/sowing window (March–May) and crop duration in the case of the Ratoon sugarcane crop. Further, we implemented the same model in an adjacent province to evaluate the performance of the proposed approach. The overall accuracy of 75.86% was obtained, and precision and recall for sugarcane were 0.79 and 0.76, respectively. This shows that strong temporal descriptors derived from an unsupervised approach can be combined with supervised learning for mapping sugarcane areas in the absence of ground labels.