From drone to satellite: Alfalfa fractional cover estimation with machine learning and sentinel-2

Alfalfa is a widely cultivated perennial forage crop with high yields and excellent quality. Accurate and timely monitoring of alfalfa growth enables the optimization of cultural practices to maximize yield. This study proposes a framework for estimating alfalfa fractional vegetation cover (FVC) using Sentinel-2 multispectral satellite imagery and machine learning (ML) algorithms. To establish precise reference data, alfalfa maps were generated from unmanned aerial vehicle (UAV) imagery using a combination of unsupervised k-nearest neighbor (kNN) and supervised random forest (RF) algorithms. The UAV images were classified into three classes: (1) alfalfa, (2) grass, and (3) dead vegetation. The classified images were then upscaled to generate FVC reference data that matched the pixels of Sentinel-2 imagery. Subsequently, two ML regression models were trained, using RF and extreme gradient boosting (XGB), to estimate alfalfa FVC. The results of this study indicated that both RF and XGB regression models could effectively estimate alfalfa FVC using Sentinel-2 data, achieving an R² value of 0.95 and a mean absolute error (MAE) of about 3.7% on the test data. However, RF slightly outperformed XGB, especially when estimating near-zero FVC values. Additionally, the results indicated that normalized difference vegetation index (NDVI) and soil-adjusted vegetation index (SAVI) were among the top four most important features in both RF and XGB. Following training, the RF model was applied to map FVC for alfalfa in the first growing cycle. The results demonstrated a complete correspondence with the RGB image visualization. Farmers and decision-makers can utilize the FVC maps generated from this study to manage alfalfa fields, identify areas with dead or damaged plants early in the spring, and make informed decisions regarding crop rotation as needed.