Identification of phenological characteristics and mapping lavender using Sentinel-2 time series data

Lavender not only has ornamental value but also significant economic and ecological benefits. Particularly in the dry areas of Northwest China, lavender plays a vital role in environmental protection, returning farmland to forest or grassland, and preventing wind erosion. Additionally, the lavender industry has become an integral part of the regional economic development in the Yili Kazakh Autonomous Prefecture, Xinjiang. Therefore, rapidly and accurately obtaining lavender planting information is vital for fostering the long-term growth of the lavender industry in the region. This study utilized high-resolution Sentinel-2 satellite imagery combined with NDVI and RENDVI₇₈₃ time series data reconstructed using the Savitzky-Golay filtering method to extract seven key phenological features of lavender. The out-of-bag (OOB) error method was used to optimally select these features, and based on this, six different classification schemes were constructed. These schemes were integrated with three types of machine learning models—Random Forest, Decision Tree, and Maximum Likelihood Classification—to explore their applicability in remote sensing identification studies of lavender. The results showed that NDVI and RENDVI₇₈₃ time series vegetation indices displayed consistent trends during the lavender growing season, with a significant increase from the seedling stage in April to the peak blooming period in early May, reaching a peak during the flower-belling period from late June to July. Although the peak values in the RENDVI₇₈₃ time series were lower, the distinction between crops was more apparent, and the phenological features were clearer. Phenological features such as Amplitude, Base, Midpoint, and Max Value scored above 0.20 in the OOB selection, thus were considered key variables and included in subsequent classification experiments. Among all classification schemes, the combination of RENDVI₇₈₃ and optimally selected phenological features (Scheme F) achieved the highest classification accuracy, where the Random Forest model (RF-F) achieved an overall accuracy of 93.21% and a Kappa coefficient of 89.98. The Decision Tree (DT-F) and Maximum Likelihood Classification (MLC-F) methods had overall accuracies of 92.41% and 90.68%, with Kappa coefficients of 88.86 and 86.02, respectively. The research results provide new insights and references for the rapid and precise extraction of planting information for crops such as lavender.