Precision forage-yield estimation in wetland rangelands through Multi-sensor satellite data fusion and machine learning

Context: In wetland meadow–rangeland ecosystems such as the Kızılırmak Delta (Türkiye), which is listed on the UNESCO World Heritage Tentative List, forage yield can shift rapidly across space and time due to hydrological fluctuations, climate variability, and grazing pressure. This dynamism complicates grazing planning and carrying-capacity assessments while maintaining a conservation–use balance. Objectives To (i) compare how Sentinel-2, Landsat-8/9, and multi-sensor fusion (Combined) approaches influence the accuracy of forage-yield prediction, and (ii) translate within-season yield dynamics into spatial products suitable for monitoring and decision support in wetland rangeland management. Methods Forage yield measured during the 2022–2023 growing seasons (June–November; n  = 411) was linked to cloud-masked surface reflectance products processed in Google Earth Engine. For Sentinel-2 (10 m), Landsat-8/9 (30 m), and Combined scenarios, predictors included NDVI, NDRE, GNDVI, and OSAVI; SWIR bands; phenological timing (day of year, DOY); and 30-day lagged variables (lag_NDVI and lag_precip from CHIRPS). Random Forest (RF), Gradient Boosting Regressor (GBR), and Extreme Gradient Boosting (XGB) models were trained using an 80% training set and a 20% independent test set, and evaluated with r , R², RMSE, nRMSE, MAE, and MBE. Monthly yield maps and pixel-wise trend analysis were further used to characterize within-season spatial variability. Key results: Model performance ranged from R² = 0.14–0.26 for Sentinel-2, R² = 0.46–0.54 for Landsat-8/9, and R² = 0.56–0.68 for the Combined scenario. The best performance was achieved by Combined + XGB (R² = 0.676; RMSE = 142.1 kg ha⁻¹; MAE = 109.7 kg ha⁻¹; nRMSE = 28.82%). Alongside SWIR and NDVI/OSAVI, DOY and lagged predictors made meaningful contributions to prediction skill. Conclusion Multi-sensor fusion coupled with boosting-based models substantially improves forage-yield prediction in wetland rangelands and enables robust spatial products for within-season monitoring. Implications and impacts: The resulting maps and trend indicators provide an actionable decision-support backbone for UNESCO-sensitive wetland landscapes, helping to guide grazing intensity, rotation/resting strategies, and adaptive management under hydrological and climatic variability while safeguarding the conservation–use balance.