Self-supervised learning for mapping rice areas reduces the need for field surveys

Traditional approaches to mapping rice areas depend on detailed field surveys to gather label data. This data gathering, while thorough, is often time-consuming, and costly. Further, for long-term rice area mapping, especially at large scales, ground validation can only be made in recent years, which may lead to uncertainties in rice areas of the past. To solve this limitation, this study introduces a novel method for classifying rice cultivation areas without the need for field data by employing a self-supervised learning framework within Google Earth Engine. Using MODIS data to calculate indices such as the Normalized Difference Vegetation Index (NDVI), the Enhance Vegetation Index (EVI), and the Land and Surface Water Index (LSWI), a pseudo-label boundary was delineated. The delineation of the boundary involved marking flooded regions, where LSWI + 0.05 ≥ NDVI EVI. Within these flooded regions, instances when EVI in the first 40 days was at least half of the peak EVI during the growing season were assigned as rice. In contrast, regions that did not meet these criteria were considered for non-rice. As proof of concept, the delineated pseudo-label boundaries of rice and non-rice were randomly sampled and trained on several machine learning models like random forests, support vector machines, gradient-boosted trees, and decision trees, to classify rice areas in Punjab, India, from 2003 to 2022. The random forest model demonstrated superior performance, achieving an Area Under the Curve of receiver-operating characteristics (AUC) of 0.71, compared to other models (AUC of ~0.55). Furthermore, comparing the self-supervised models against the same machine learning models, which were traditionally trained on field survey data (228 ground points: 164 rice, 64 non-rice was collected), the self-supervised models showed ~10% higher performance than their traditionally supervised counterparts. Therefore, this study demonstrates that using this self-supervised modeling framework reduces the need for field-based annotations, while still providing reasonably accurate rice area maps.