نوع مقاله : کاربردی

نویسندگان

1 دانشجوی دکتری گروه علوم خاک، دانشکده کشاورزی، واحد اصفهان (خوراسگان)، دانشگاه آزاد اسلامی، اصفهان، ایران.

2 دانشیار گروه علوم خاک، دانشکده کشاورزی، واحد اصفهان (خوراسگان)، دانشگاه آزاد اسلامی، اصفهان، ایران.

3 استاد گروه علوم خاک، دانشکده کشاورزی، دانشگاه صنعتی اصفهان، اصفهان، ایران.

4 دانشیار گروه علوم خاک، موسسه گاند برای محیط زیست، دانشگاه ورمانت، خیابان کولچستر 210، برلینگتون، VT 05401، آمریکا.

10.22055/agen.2025.48843.1761

چکیده

ویژگی‌های فیزیکوشیمیایی و مکانیکی خاک در سلامت خاک، بهره‌وری کشاورزی و ارزیابی پایداری آن نقش اساسی دارند. درک توزیع مکانی این ویژگی‌ها برای مدیریت مؤثر اراضی و کشاورزی پایدار ضروری است. نقشه‌برداری رقومی خاک (DSM) با استفاده از داده‌های محیطی و الگوریتم‌های یادگیری ماشین (ML)، ابزاری قدرتمند برای پیش‌بینی ویژگی‌های خاک است. هدف این پژوهش، پیش‌بینی توزیع مکانی میانگین وزنی قطر خاکدانه‌ها (MWD) و کربن آلی خاک (SOC) با استفاده از روش DSM و رتبه‌بندی متغیرهای محیطی مؤثر با مدل‌هایML و مقایسه این مدل‌ها ‌در منطقه فریدن، استان اصفهان بود. شش الگوریتم ML شامل شبکه‌ عصبی مصنوعی (ANN)، جنگل تصادفی (RF)، مدل کوبیست (Cubist)، ماشین بردار پشتیبان (SVM)، الگوریتم K نزدیک‌ترین همسایه (KNN) و مدل رگرسیون درختی توسعه‌یافته (BRT) برای ارتباط دادن متغیرهای محیطی و ویژگی‌های خاک استفاده شدند. مدل‌سازی با چهار مجموعه‌داده شامل داده‌های سنجش از دور، توپوگرافی، ویژگی‌های خاک و داده‌های طبقه‌بندی شده انجام شد و عملکرد مدل‌ها با استفاده از ضریب تعیین (R²)، ریشه میانگین مربعات خطا (RMSE) و میانگین خطای مطلق (MAE) ارزیابی گردید. نتایج نشان داد که بهترین دقت پیش‌بینی برای MWD، در مدل KNN در سناریوی چهارم (59/0=R²، 19/0= RMSEو 16/0=MAE) و برای SOC، در مدل cubist در سناریوی چهارم (78/0=R²، 12/4= RMSEو 53/3=MAE)، به‌دست آمد. سیلت (Silt) مهم‌ترین متغیر برای MWD بود و VV و CHND در رتبه‌های بعدی قرار گرفتند. برای پیش‌بینی SOC، MRVBF و Clay بیشترین تأثیر را داشتند. در نهایت، مدل‌های cubist و KNN با ترکیب داده‌های توپوگرافی و اقلیمی، به‌عنوان ابزاری مؤثر برای پیش‌بینی ویژگی‌های کیفی خاک و تولید نقشه‌های رقومی خاک انتخاب شدند. این روش‌ها به‌ویژه در مناطقی با اقلیم مشابه و شرایط نمونه‌برداری محدود، می‌توانند به بهبود بهره‌وری کشاورزی و ارتقای مدیریت پایدار اراضی کمک کرده و همچنین به سیاست‌گذاران در شناسایی مناطق بحرانی و کاهش اثرات منفی زیست‌محیطی یاری رسانند.

کلیدواژه‌ها

موضوعات

عنوان مقاله [English]

Digital mapping of soil physical and chemical properties using some machine learning algorithms and environmental variables in the Fereydan region, Isfahan Province

نویسندگان [English]

  • Nazanin Sadat Emami 1
  • Elham Chavoshi 2
  • Shamsollah Ayoubi 3
  • Naser Honarjoo 2
  • Mojtaba Zeraatpisheh 4

1 Ph.D. student, Department of Soil Science, College of Agriculture, Isfahan (Khorasgan) Branch, Islamic Azad University, Isfahan, Iran.

2 Associate Professor, Department of Soil Science, College of Agriculture, Isfahan (Khorasgan) Branch, Islamic Azad University, Isfahan, Iran.

3 Professor, Department of Soil Science, College of Agriculture, Isfahan University of Technology, Isfahan, 841568311, Isfahan, Iran.

4 Associate Professor, Department of Soil Science, Gund Institute for Environment, University of Vermont, 210 Colchester Ave, Burlington, VT 05401, USA.

چکیده [English]

Introduction

Soil physical and chemical properties play critical roles in soil health, agricultural productivity, and natural resource management. These properties are key indicators for assessing soil sustainability and carbon storage. Understanding their spatial distribution is essential for effective land management and resource modeling, particularly within the context of sustainable agriculture. The mean weight diameter of soil aggregates (MWD) is a crucial functional parameter that significantly influences the carbon cycle, soil structural stability, and agricultural productivity. Similarly, soil organic carbon (SOC) plays a vital role in environmental sustainability. Digital soil mapping (DSM), which integrates data from environmental covariates and machine learning (ML) algorithms, is a powerful tool for predicting soil properties and enhancing sustainable land management practices. The DSM techniques have proven highly effective in estimating soil parameters across diverse landscapes, minimizing the need for extensive field surveys and enabling data-driven decision-making in precision agriculture. The primary objective of this study was to predict the spatial distribution of MWD and SOC using DSM. Additionally, this study aimed to rank the most influential environmental variables affecting these properties through ML-based modeling and compare the performance of various ML methods.

Materials and Methods This study was conducted in the Fereydan region (50º 6' E, 32º 56' N), Isfahan Province, Iran. A total of 100 surface soil samples (0 to 20 cm depth) were collected from various land uses, including irrigated croplands, dryland farms, and pastures, using a systematic sampling network. The collected samples were air-dried, sieved through a 2 mm mesh, and then subjected to laboratory analyses. Six ML algorithms—artificial neural network (ANN), random forest (RF), cubist model, support vector machines (SVM), k-nearest neighbors (KNN), and boosted regression trees (BRT)—were employed to relate environmental variables to the soil properties under investigation. For the modeling process, four datasets were used as predictor environmental covariates: remote sensing data, topography derived from a digital elevation model (DEM), soil properties, and classified land data. The models' accuracy was evaluated across four different scenarios, each using one or more of these datasets. The most suitable environmental variables were selected through a multicollinearity test using the variance inflation factor (VIF). Finally, to model the two soil properties, the dataset was divided into calibration (70%) and validation (30%) subsets. The models' performance was evaluated using various statistical metrics, including the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE).

Results and Discussion The results of model validation revealed that the highest prediction accuracy for MWD was achieved using the KNN model in the fourth scenario (R²=0.59, RMSE=0.19, MAE=0.16). These values indicate a moderate yet significant ability of the KNN model to predict MWD. In contrast, for SOC, the cubist model in the fourth scenario yielded the highest accuracy (R²=0.78, RMSE=4.12, MAE=3.53). These results suggest that the cubist model was particularly effective in capturing the complex relationships between environmental variables and SOC, providing a robust framework for predicting soil carbon content. These findings highlight the strong potential of ML models to capture the intricate interactions between soil properties and environmental factors, allowing for precise spatial predictions of soil characteristics. The ability of these models to process large volumes of data and uncover hidden patterns is particularly useful in soil science, where the relationship between soil properties and environmental variables can be complex and nonlinear. The results also revealed that different environmental variables played varying roles in explaining the spatial variability of MWD and SOC. For MWD prediction, silt content was found to be the most influential variable, followed by VV and CHND, both of which were also significant factors affecting the spatial distribution of MWD. In contrast, for SOC prediction, the most important variables were the multi-resolution valley bottom flatness index (MRVBF) and clay content. These factors are critical for understanding the storage and cycling of organic carbon in soil, as both moisture retention and clay particles are known to influence the stabilization of soil organic matter. The relationship between these variables and SOC highlights the importance of considering soil texture and moisture dynamics when predicting soil carbon stocks.

Conclusion The results of this study demonstrated that ML models, such as cubist and KNN can effectively predict soil quality properties across different regions, particularly when combined with topographic and climatic data. These approaches offer an efficient and rapid method for generating digital soil maps and supporting sustainable land management. By leveraging multiple environmental datasets, ML-based DSM enhances soil modeling accuracy while reducing reliance on extensive soil sampling. These models are especially useful in areas with similar climates where full-scale soil sampling may not be feasible, providing a valuable tool for estimating soil quality attributes and enhancing agricultural productivity. Ultimately, the findings of this research can help policymakers and farmers identify critical areas and adopt optimal strategies to improve soil quality and mitigate adverse environmental impacts. To further improve the accuracy and stability of predictions, future studies should incorporate high-resolution spatiotemporal data, advanced algorithms such as cubist and KNN, and dynamic models for analyzing temporal variations. Additionally, investigating soil–water–plant interactions and assessing the impacts of climate change could play a crucial role in enhancing agricultural productivity and promoting the sustainable management of soil resources.

کلیدواژه‌ها [English]

  • Digital soil mapping
  • Machine learning
  • Environmental variables
  • Soil properties
  • Modeling accuracy.Feature selection