A UAV-to-Satellite Scaling Framework for Monitoring Cotton Boll Opening Using Sentinel-2 Earth Observations

Cotton boll opening is an important late-season indicator for maturity assessment, defoliation timing, and harvest planning, but field-level monitoring remains challenging because visible lint progression varies spatially and temporally. This study evaluates whether UAV-derived cotton visible lint percentage, aggregated at the Sentinel-2 10 m grid scale, can be estimated using Sentinel-2 spectral bands, vegetation indices (VIs), and accumulated growing degree days (AGDD) as phenological predictors. UAV multispectral imagery was used to derive visible lint percentage through red-band thresholding and segmentation within canopy masks. The UAV-derived visible lint information was summarized within fixed Sentinel-2 10 m grid cells to generate satellite-compatible response labels. Four supervised regression models, eXtreme Gradient Boosting (XGBoost), Random Forest (RF), k-Nearest Neighbors (kNN), and Neural Network/Multilayer Perceptron (NNET/MLP), were evaluated using raw and transformed target formulations. Raw visible lint percentage produced relatively high explanatory power for tree-based models, with R2 values of 0.73 for both XGBoost and RF. However, the target distribution was strongly right-skewed and dominated by low visible lint values, with a mean PCTOPEN of 4.05%Open, motivating the evaluation of target transformations to reduce target skewness while assessing their impact on predictive performance. On the original PCTOPEN scale, the raw target produced RMSE = 4.38%Open points, MAE = 2.31%Open points, and MedianAE = 0.68%Open points. The square-root transformation provided the strongest overall predictive performance, maintaining R2 = 0.73 and RMSE = 4.38%Open points while reducing MAE to 2.14%Open points and MedianAE to 0.41%Open points. Stronger transformations further reduced the typical prediction errors, with MedianAE = 0.35, 0.33, and 0.31%Open points for the cube-root, fourth-root, and fifth-root transformations, respectively. However, these improvements were accompanied by progressively lower R2 values (0.72, 0.71, and 0.69) and higher RMSE values (4.48, 4.58, and 4.68%Open points), indicating that stronger transformations reduced typical prediction errors at the expense of overall predictive performance. These results indicate that UAV-derived visible lint percentage can be linked with Sentinel-2 observations for satellite-scale regression modeling, but prediction uncertainty remains influenced by target skewness, mixed 10 m pixels, canopy obstruction, and limited UAV acquisition density. Additional UAV–Sentinel-2-aligned acquisitions, supporting multi-field observations, and broader validation across fields, seasons, cultivars, and production environments are needed to improve robustness and support operational cotton boll-opening tracking applications.