نوع مقاله : مقاله پژوهشی
نویسندگان
1 دانشجوی دکتری علوم و مهندسی مرتع، دانشکده منابع طبیعی و علوم زمین، دانشگاه شهرکرد، شهرکرد، ایران
2 استادیار گروه مهندسی طبیعت، دانشکده منابع طبیعی و علوم زمین، دانشگاه شهرکرد، شهرکرد، ایران
3 دانشیار گروه مهندسی طبیعت، دانشکده منابع طبیعی و علوم زمین، دانشگاه شهرکرد، شهرکرد، ایران
4 استادیار گروه احیاء مناطق خشک و کوهستانی، دانشکده منابع طبیعی، دانشگاه تهران، کرج، ایران
چکیده
مقدمه
نقشههای کاربری/پوشش زمین از ابزارهای کلیدی در مدیریت منابع طبیعی، برنامهریزی منطقهای و دستیابی به توسعه پایدار به شمار میآیند و ازاینرو ضرورت پایش دقیق و بهروز آنها بیشازپیش احساس میشود. تغییرات ناشی از عوامل انسانی و طبیعی، الگوهای کاربری زمین را بهطور مستمر دستخوش تحول کرده و پیامدهای قابلتوجهی بر بومسازگانهای محلی و جهانی ایجاد کردهاند. در دهههای اخیر، پیشرفت فناوری سنجش از دور و الگوریتمهای یادگیری ماشین تحول مهمی در استخراج و طبقهبندی دادههای مکانی ایجاد کرده است. سامانه گوگل ارث انجین (GEE) بهعنوان بستری قدرتمند برای پردازش کلاندادههای مکانی-زمانی، امکان تولید نقشههای دقیق و بهروز را فراهم میکند. در این چارچوب، مطالعه حاضر با تمرکز بر ترکیب تصاویر چندزمانه، بهرهگیری از دادههای کمکی و مقایسه سه الگوریتم یادگیری ماشین در یک حوزه آبخیز وسیع و ناهمگن همچون کارون ۱، با هدف بهبود دقت طبقهبندی و تقویت قابلیت پایش تغییرات بلندمدت پوشش زمین انجام شده است.
مواد و روشها
بهمنظور ارزیابی تغییرات کاربری و پوشش زمین در حوزه آبخیز کارون ۱، تصاویر سنجندههای ETM+ لندست ۷ (سال ۲۰۰۲) وOLI لندست ۸ (سال ۲۰۲۴) با پوشش ابری کمتر از ۱۰ درصد و با لحاظ میانگین بارندگی بلندمدت، از طریق GEE و در قالب دادههای بازتاب سطحی تصحیحشده مورد استفاده قرار گرفت. تصاویر ترکیبی حاصل از نُه صحنه تصویری لندست طی دوره اوج رشد پوشش گیاهی (اردیبهشت تا تیر ماه) با فیلتر میانه تهیه و سپس مطابق مرز حوضه برش داده شدند. در ادامه، تعداد ۱۹۲۰ نقطه آموزشی در قالب هفت کلاس کاربری/پوشش زمین مطابق طبقهبندی آندرسون و بر پایه دادههای میدانی، تصاویر هوایی و گوگل ارث برداشت شد. این دادهها بهصورت تصادفی به دو مجموعه آموزش (۶۰ درصد) و ارزیابی (۴۰ درصد) تقسیم شدند. متغیرهای کمکی (مانند NDVI، NDBI، NDWI و DSM) استخراج و در کنار باندهای طیفی اصلی گنجانده شدند. آزمایشهای طبقهبندی در GEE با استفاده از سه الگوریتم نظارت شده CART، RF و SVM اجرا شدند. پارامترهای مدل و رویههای آموزشی برای اطمینان از تکرارپذیری و سازگاری در بین روشها پیکربندی شدند.
نتایج و بحث
نتایج نشان داد که الگوریتمهای CART،RF و SVM نقشههای طبقهبندی را با دقت عالی تولید کردند و بهرهگیری از شاخصهای پوشش گیاهی و دادههای کمکی، دقت کلی و ضریب کاپای طبقهبندی را در هر دو سال افزایش داد. بیشینه دقت کلی و کاپا مربوط به الگوریتم SVM بود که در سال ۲۰۰۲ بهترتیب ۹۳ و 91.5 درصد و در سال ۲۰۲۴ برابر ۹۳ و ۹۲ درصد بهدست آمد. همچنین بر اساس نتایج هر سه الگوریتم، بخش عمدهای از مساحت حوضه (بهطور متوسط ۴۰ درصد) از مراتع و حدود ۲۷ درصد از جنگل تشکیل شده است. روند زمانی نتایج نیز نشاندهنده کاهش مساحت مراتع و جنگلها و بهویژه کاهش پهنههای آبی بر اساس الگوریتم SVM است.
نتیجهگیری
نتایج این پژوهش نشان داد که بهکارگیری الگوریتمهای یادگیری ماشین در بستر گوگل ارث انجین امکان تولید نقشههای دقیق کاربری/پوشش زمین و پایش مؤثر تغییرات محیطی را در یک حوزه آبخیز وسیع و ناهمگن فراهم میسازد. نتایج حاصل میتواند بهعنوان ابزاری کارآمد در برنامهریزی کاربری اراضی، مدیریت منابع طبیعی، پایش تخریب پوشش گیاهی و کنترل تغییرات کاربری در منطقه مورد استفاده قرار گیرد. با این حال، محدودیتهای ناشی از وضوح مکانی تصاویر لندست و خطاهای اختلاط پیکسلی بهویژه در مرز بین طبقات کاربری/پوشش زمین ازجمله چالشهای این مطالعه محسوب میشوند. بر این اساس، پیشنهاد میشود در پژوهشهای آتی از دادههای با قدرت تفکیک مکانی بالاتر مانند تصاویر سنتینل برای بهبود دقت طبقهبندی بهویژه در پوششهای گیاهی بهره گرفته شود.
کلیدواژهها
عنوان مقاله [English]
Monitoring and detection of land use/land cover change using machine learning algorithms on the google earth engine platform: a case study of the Karun 1 Watershed
نویسندگان [English]
- Sina Nabizadeh 1
- Ali asghar Naghipour 2
- Ataollah Ebrahimi 3
- Hamidreza Keshtkar 4
- Elham Ghehsareh 3
1 Ph.D. Student, Department of Nature Engineering, Faculty of Natural Resources and Earth Sciences, Shahrekord University, Shahrekord, Iran
2 Assistant Professor, Department of Nature Engineering, Faculty of Natural Resources and Earth Sciences, Shahrekord University, Shahrekord, Iran
3 Associate Professor, Department of Nature Engineering, Faculty of Natural Resources and Earth Sciences, Shahrekord University, Shahrekord, Iran
4 Assistant Professor, Department of Arid and Mountainous Regions Reclamation, Faculty of Natural Resources, University of Tehran, Iran, Karaj, Iran
چکیده [English]
Introduction
Land use/land cover (LULC) maps are among the key tools for natural resource management, regional planning, and achieving sustainable development; therefore, the need for their accurate and up-to-date monitoring is increasingly emphasized. Continuous changes in land use driven by natural and human-induced factors have significantly affected local and global ecosystems. In recent decades, advances in remote sensing technologies and machine learning algorithms have led to major improvements in the extraction and classification of spatial data. The Google Earth Engine (GEE) platform, as a powerful cloud-based infrastructure for processing large-scale spatiotemporal data, provides an efficient framework for producing accurate and updated maps. Within this context, the present study focuses on the integration of multi-temporal satellite images, the use of auxiliary data, and the comparison of three machine learning algorithms over a large and heterogeneous watershed (Karun 1), aiming to improve classification accuracy and enhance the capability for long-term monitoring of LULC changes.
Materials and methods
To assess LULC changes in the Karun 1 watershed, Landsat 7 ETM+ (2002) and Landsat 8 OLI (2024) images with cloud cover less than 10% and considering long-term mean precipitation were retrieved and processed as surface reflectance products in the GEE platform. Composite images were generated from nine Landsat scenes during the peak growing season (May to July) using a median filter and were then clipped to the watershed boundary. A total of 1,920 training samples representing seven LULC classes based on the Anderson classification scheme were collected using field survey data, aerial photographs, and Google Earth imagery. The reference dataset was randomly split into training (60%) and evaluation (40%) subsets Auxiliary variables (such as NDVI, NDBI, NDWI, and a DSM) were derived and included alongside original spectral bands. Classification experiments were implemented in GEE using three supervised algorithms: CART, RF, and SVM. Model hyper parameters and training procedures were configured to ensure reproducibility and consistency across methods.
Results and discussion
The results showed that the CART, RF, and SVM algorithms produced classified maps with excellent accuracy. The incorporation of vegetation indices and auxiliary data improved both the overall accuracy and the Kappa coefficient for both study years. The highest overall accuracy and Kappa coefficient were achieved by the SVM algorithm, with values of 93% and 91.5% in 2002, and 93% and 92% in 2024, respectively. According to the results of all three algorithms, rangelands constitute the largest proportion of the watershed area (on average about 40%), followed by forests (approximately 27%). The temporal analysis indicated a decreasing trend in the area of rangelands and forests, as well as a notable reduction in water bodies, particularly based on the SVM results.
Conclusions
The results of this study demonstrated that applying machine learning algorithms within the Google Earth Engine platform enables the production of accurate land use/land cover maps and the effective monitoring of environmental changes over a large and heterogeneous watershed. The obtained results can be used as an efficient tool for land use planning, natural resource management, monitoring vegetation degradation, and controlling land use changes in the study area. However, limitations related to the spatial resolution of Landsat imagery and pixel-mixing errors, particularly along the boundaries between LULC classes, are considered among the main challenges of this study. Accordingly, it is recommended that future research utilize higher spatial resolution data, such as Sentinel imagery, to improve classification accuracy, especially for vegetation cover mapping.
کلیدواژهها [English]
- Auxiliary data
- Classification and Regression Tree (CART)
- Random Forest (RF)
- Remote sensing
- Support Vector Machine (SVM)