Evaluation of land suitability for wheat cultivation using different machine learning algorithms and GIS techniques: Gözlü agricultural farm example

Today, food security is a significant concern for decision-makers, particularly in regions with arid and semi-arid climates and limited natural resources. In this context, many developed and developing countries are designing crop-specific land evaluation models within the framework of precision agriculture to ensure the optimal use of soil and water resources. Among these models, machine learning algorithms are increasingly being employed as fast and reliable decision-support tools, especially for crops of strategic importance. In fact, due to its strategic role in human nutrition, assessing land suitability for wheat production has become a key focus for both researchers and policymakers. Given the strategic importance of wheat cultivation in Turkey, region-specific studies hold particular value. However, a crop-specific land evaluation approach supported by machine learning algorithms has not yet been applied to wheat cultivation in Turkey. This study presents a preliminary effort to generate land suitability maps for wheat cultivation and identify key cultivation parameters using machine learning algorithms integrated with Geographic Information Systems, tailored to the arid climate conditions of the Central Anatolia region. In this study, 0–30 cm deep soil samples were collected from 119 different locations using the soil map of the Konya-Gözlü Agricultural Enterprise, which represents the general soil characteristics and climatic conditions of the region. Topographic indicators (soil depth, slope, and gravel), physical indicators (soil texture, bulk density, and available water capacity), chemical indicators (soil pH, electrical conductivity, calcium carbonate content, and organic matter), and the long-term normalized difference vegetation index (LT-NDVI) were selected as input parameters. These parameters were modeled using five different machine learning algorithms artificial neural networks, k-nearest neighbors, support vector machines, random forests, and the XgbTree algorithm to generate land suitability maps for wheat cultivation. To construct the machine learning dataset, 238 randomly selected points within the Gözlü Agricultural Enterprise were classified into two sets: 70% (167 points) for training and 30% (71 points) for testing. Following spatial interpolation in the ArcGIS Pro environment, each pixel of the rasters representing soil parameters and long-term NDVI was assigned to a specific land suitability class for wheat cultivation using the R programming environment. The analysis of parameter importance revealed that soil texture and available water capacity were the most influential factors in predicting land vii suitability maps when using artificial neural networks, support vector machines, and the XgbTree algorithm. In contrast, for the random forest and k-nearest neighbors algorithms, soil depth and soil pH emerged as the most critical parameters in determining land suitability classes. Overall, the results emphasized that soil texture, available water capacity, and soil depth had the greatest impact on the generation of land suitability maps. These were followed by moderately influential parameters, including chemical properties such as calcium carbonate content and electrical conductivity, as well as the topographic indicator soil slope. Conversely, parameters such as bulk density, organic matter, and long-term NDVI were found to have minimal importance in the prediction process. The performance evaluation of the machine learning algorithms showed that the XgbTree algorithm outperformed the others, achieving 98% accuracy and a kappa coefficient of 0.81. The random forest algorithm ranked second, yielding strong results with 94% accuracy and a kappa coefficient of 0.78. In contrast, the artificial neural networks and k-nearest neighbors algorithms demonstrated relatively lower predictive performance, achieving 75% and 80% accuracy and kappa coefficients of 0.56 and 0.68, respectively. The support vector machines algorithm exhibited the lowest performance, with 71% accuracy and a kappa coefficient of 0.49. These findings underscore the superior predictive performance of the XgbTree and random forest algorithms. According to the Area Under the Curve (AUC) analysis conducted on the validation dataset, the XgbTree algorithm achieved an AUC value of 0.90. Similarly, the random forest algorithm also demonstrated strong performance with an AUC value of 0.90. The k-nearest neighbors algorithm exhibited a comparable level of performance, reaching an AUC value of 0.91. In contrast, the artificial neural networks algorithm showed lower performance with an AUC value of 0.81, while the support vector machines algorithm recorded the lowest performance with an AUC value of 0.75. Based on the XgbTree and random forest algorithms both of which achieved the highest model accuracy the spatial distribution of wheat suitability classes across the agricultural lands of the enterprise was determined to be similar: S1 (highly suitable) comprised 27.90%–25.23%, S2 (moderately suitable) 11.74%–19.68%, S3 (marginally suitable) 24.34%–27.66%, and N (not suitable) 36.00%–27.40%. The results provide preliminary insights to support decision-making processes in land-use planning. The study determined that machine learning models hold significant potential in crop-specific land evaluation studies and may serve as a guide for future research on applying machine learning algorithms in suitability assessments. Although the XgbTree algorithm demonstrated the high accuracy on validation data, the artificial neural networks algorithm showed the highest performance during the training phase. These findings indicate that evaluation processes should not rely solely on a single model but should incorporate different soil and land factors. Further research is recommended to enhance interpretation by comparing various parameters used in the modeling process. Keywords: GIS, machine learning, Middle Anatolia, suitability classification, wheat.