پیش‌بینی و نقشه‌برداری خصوصیات فیزیکی و شیمیایی خاک در مناطق نیمه‌خشک: تحلیل مقایسه‌ای الگوریتم‌های یادگیری ماشین

درک تغییرات مکانی ویژگی‌های خاک برای نظارت بر قابلیت‌های زمین و مدیریت پایدار منابع خاک بسیار حائز اهمیت است. این مطالعه با هدف ارزیابی دقت و کارایی مدل‌های یادگیری ماشین در برآورد خصوصیات فیزیکی و شیمیایی خاک انجام شده است. برای این منظور، تعداد 110 نمونه خاک از دو عمق 0-30 سانتی‌متر و 30-60 سانتی‌متر جمع‌آوری و مورد بررسی قرار گرفت. مدل‌های یادگیری ماشین مورد استفاده شامل rf، cubist و rf-cubist بودند و داده‌ها به دو بخش آموزش (70٪) و آزمون (30٪) تقسیم شدند. نتایج نشان داد که برای عمق 0-30 سانتی‌متر، مدل rf برای شن با rmse، bias، cor و rrse به ترتیب معادل 10.19درصد، 0.30-، 0.92 و 0.41 بهترین عملکرد را داشت. همچنین، مدل cubist برای کربن آلی با rmse، bias، cor و rrse به ترتیب معادل 0.23درصد، 0.005، 0.71 و 0.70 بهترین برآورد را داشته است. مدل rf-cubist نیز در پیش‌بینی سیلت با rmse، bias، cor و rrse به ترتیب معادل 7.80درصد، 0.87-، 0.87 و 0.28 نتایج قابل قبولی به‌دست آورد. در عمق 30-60 سانتی‌متر، مدل rf برای شن با rmse، bias، cor و rrse به ترتیب معادل 11.42درصد، 0.23-، 0.88 و 0.47 بهترین عملکرد را داشت، و مدل cubist برای کربن آلی با rmse، bias، cor و rrse به ترتیب معادل 0.26درصد، 0.05، 0.53 و 0.85 بهترین بود. همچنین، مدل rf-cubist در پیش‌بینی سیلت با rmse، bias، cor و rrse به ترتیب معادل 9.28درصد،0.58-، 0.83 و 0.30 عملکرد خوبی داشت. نقشه‌های عدم قطعیت برای پارامترهای مختلف نیز نشان دادند که مدل rf در بسیاری از نقاط دارای عدم قطعیت کمتری نسبت به سایر مدل‌ها است. این پژوهش نشان می‌دهد که مدل‌های یادگیری ماشین، به‌ویژه rf-cubist و cubist، ابزارهای موثری برای پیش‌بینی خصوصیات خاک بوده و می‌توانند به مدیریت بهتر منابع خاک و کشاورزی پایدار کمک کنند. 

prediction and mapping physical and chemical soil properties in semi-arid regions: a comparative analysis of machine learning algorithms

1- introductionin recent years, with the rise of challenges such as food security, land degradation, water resource scarcity, climate change, and biodiversity conservation, the importance of understanding the physical and chemical properties of soils has increased significantly. these properties have a direct impact on food production, carbon storage, water filtration, and climate regulation, making accurate information on these characteristics essential for environmental and agricultural management. traditional methods of measuring soil properties have limited applicability over large areas due to their cost and time requirements. however, the emergence of machine learning algorithms and remote sensing data has facilitated more precise mapping. digital soil mapping (dsm) has emerged as an innovative approach that combines advanced algorithms with environmental variables to provide high-quality estimates of soil properties. recent studies indicate that utilizing models such as random forest (rf) and cubist, paired with topographic and remote sensing data, is effective for predicting soil variables and can significantly enhance the management of natural resources, agriculture, and watersheds. in this study, sampling locations were determined using the cubist latin hypercube sampling (clhs) method.2- material and methods auxiliary variables in the latin hypercube model included bands and indices from sentinel-2 satellite imagery, encompassing visible spectrum bands, short-wave infrared, vegetation indices, and properties derived from a digital elevation model (dem), including elevation, slope, aspect, and valley flatness index with high resolution. the boruta method was employed to select influential variables. the machine learning models used in this study included rf, cubist, and rf-cubist, with the data divided into two sections: training (70%) and testing (30%). accordingly, all data were evaluated as testing data based on the aforementioned method.3- results the results indicated that for the depth of 0-30 cm, the random forest (rf) model demonstrated the best performance for sand, with rmse, bias, correlation (cor), and relative root squared error (rrse) values of 10.19%, -0.30, 0.92, and 0.41, respectively. additionally, the cubist model provided the best estimates for organic carbon, with rmse, bias, cor, and rrse values of %0.23, 0.005, 0.71, and 0.70, respectively. the rf-cubist model achieved satisfactory results in predicting silt, with rmse, bias, cor, and rrse values of 7.80, -0.87, 0.87, and 0.28, respectively. at the depth of 30-60 cm, the rf model again performed best for sand, yielding rmse, bias, cor, and rrse values of %11.42, -0.23, 0.88, and 0.47, respectively. the cubist model was noted for its performance in estimating organic carbon, achieving rmse, bias, cor, and rrse values of %0.26, 0.05, 0.53, and 0.85, respectively. furthermore, the rf-cubist model yielded good results in predicting silt, with rmse, bias, cor, and rrse values of %9.28, -0.58, 0.83, and 0.30, respectively. uncertainty maps for the various parameters indicated that the rf model exhibited lower uncertainty at many locations compared to other models.4- discussion & conclusionsa specific model should provide the best results with the least error. for instance, at a depth of 0-30 cm, the random forest (rf) model exhibited higher accuracy in estimating soil physical properties, characterized by a high r² and lower rmse, compared to the cubist and rf-cubist models. conversely, the cubist model demonstrated superior accuracy in estimating soil chemical properties. at a depth of 30-60 cm, the rf-cubist model outperformed the rf and cubist models, achieving the best performance with a high r² and lower rmse. according to results from other studies, the variability of soil properties is attributed to environmental factors, diversity of parent materials, and eomorphological processes, leading to the formation of different landforms and influencing soil particles, especially clay. the low average soil organic carbon at both surface and subsurface depths (0.36 and 0.33) is affected by factors such as poor vegetation cover, limited use of organic fertilizers, and reduced soil fertility. the use of the boruta model for selecting environmental variables has reduced prediction errors compared to other methods, and variables derived from the digital elevation model (dem) have been recognized as effective in class modeling and soil property estimation. in comparing predictive models, the rf model demonstrated superior performance due to its lower error values (rmse and bias) and higher correlation coefficients compared to the cubist and rf-cubist models. nevertheless, the cubist model also displayed high accuracy in modeling due to its ability to capture the complexities between soil characteristics. multiple studies have further confirmed that the rf model provides reliable results with low uncertainty when estimating properties such as organic carbon percentage and clay content. ultimately, to improve results in future studies, it is recommended that model selection be tailored to the sampling depth and type of soil characteristics, data quality be enhanced, and the use of hybrid models be further developed.