Yazar "Kilic, M." seçeneğine göre listele

Listeleniyor 1 - 3 / 3
  • Küçük Resim Yok
    Öğe
    Data-Scarce Dryland SOC Mapping Using Landsat Spectral Indices and a Hybrid Gradient Boosting Kriging Framework
    (Pleiades Publishing Ltd, 2026) Gunal, H.; Ismael, A. M.; Kilic, M.; Budak, M.; Polat, K.; Acir, N.
    Accurate estimation of soil organic carbon (SOC) is critical in arid, data-limited regions where land degradation and climatic stress disrupt carbon cycling. We developed a hybrid learning-geostatistical framework that couples Bayesian-tuned gradient boosting decision trees (GBDT) with ordinary kriging (OK) of model residuals to capture both covariate-driven variance and residual spatial autocorrelation. Field-measured SOC observations were integrated with Landsat-8/9 spectral derivatives (vegetation and soil reflectance indices). GBDT modeled nonlinear SOC-environment relationships; OK then interpolated the spatially structured residuals, and the two components were summed to form final predictions. The hybrid raised held-out performance to R2 = 0.72 and reduced RMSE by 32.8% relative to GBDT alone. Vegetation indices explained 57% of the variance attributed to predictors (rising to 66.6% when the brightness index is included), indicating a central role for plant-soil feedbacks in carbon accumulation. The approach is readily scalable for semi-arid landscapes and supports evidence-based restoration targeting, climate adaptation planning, and MRV (monitoring, reporting, verification) of carbon outcomes.
  • Küçük Resim Yok
    Öğe
    Modeling Soil Quality Parameters in Solhan Plain Using Neural Networks and Comparing Different Algorithms for Prediction Accuracy
    (Pleiades Publishing Ltd, 2025) Alp, V.; Aydemir, S.; Bilgili, A. V.; Demirkiran, A. R.; Kilic, M.
    This study focused on modeling specific soil properties in the Bing & ouml;l Solhan Plain using soil samples collected from 85 coordinates at 0-30 cm depth, with 300 x 300 m intervals. Artificial intelligence techniques, implemented with ArcGIS 10.8 software, were employed to estimate soil characteristics efficiently and cost-effectively. The input data for the artificial neural network models consisted of electrical conductivity (EC) and clay content, while the output parameters included exchangeable cation percentage, cation exchange capacity, field capacity, and wilting point. Soil characteristics were determined through physical, chemical, and biological laboratory tests. The training process involved three algorithms: Levenberg-Marquardt (LM), scaled conjugate gradient (SCG), and bayesian regularization (BR). To enhance model performance, hyperparameter optimization was applied, with the number of iterations and network structures (comprising double hidden layers with 0-30 neurons per layer) adjusted. The training dataset was divided into 70% training, 15% testing, and 15% validation. Among the algorithms tested, the Levenberg-Marquardt algorithm yielded the best performance across all estimations. It achieved 94% accuracy with mean absolute error (MAE) and root mean squared error (RMSE) values of 0.04 and 0.046, respectively, for exchangeable cation percentage prediction. For cation exchange capacity estimation, it attained 84% accuracy, with MAE and RMSE values of 0.081 and 0.091, respectively. In field capacity estimation, it achieved 79% accuracy, with MAE and RMSE values of 0.095 and 0.118, respectively. For wilting point estimation, it reached 86% accuracy, with MAE and RMSE values of 0.057 and 0.067 respectively. The results demonstrate that the Levenberg-Marquardt algorithm consistently produced the highest R2 values and the lowest MAE and RMSE across all estimation models. This study highlights the potential of artificial intelligence techniques, particularly the Levenberg-Marquardt algorithm, to accurately model complex soil parameters that are traditionally time-consuming and costly to analyze.
  • Küçük Resim Yok
    Öğe
    Monitoring Soil Salinity in the Harran Plain: A Comparative Analysis of Machine Learning Algorithms Using Two Different Scenarios with Sentinel-2 Data
    (Pleiades Publishing Ltd, 2026) Kaplan, F.; Bilgili, A. V.; Kilic, M.
    This study aimed to predict soil salinity in the Harran Plain using remote sensing and machine learning methods. The performance of five machine learning algorithms (Random Forest (RF), Extreme Gradient Boosting (XGBoost), Support Vector Regression (SVR), Gradient Boost Machine (GBM), Classification and Regression Trees (CART) was compared under two different data scenarios: one using all parameters and another using only Sentinel-2 spectral indices. The results showed that the Random Forest (RF) algorithm achieved the highest performance in both scenarios. In the scenario using all parameters, the RF model achieved values of R2 = 0.87, RMSE = 0.154 dS/m, and MAE = 0.112 dS/m, while in the scenario using only Sentinel-2 indices, it achieved R2 = 0.83, RMSE = 0.175 dS/m, and MAE = 0.128 dS/m. SHAP analyses supported the feasibility of the model predictions and confirmed the consistency of RF. These findings demonstrate that Sentinel-2-based models provide a low-cost and effective alternative to traditional methods for large-scale areas.