تهیه نقشه نوع محصول کشاورزی از سری زمانی تصاویر لندست-8 با استفاده از روش های یادگیری ماشین (مطالعه موردی: مرودشت استان فارس) (مقاله علمی وزارت علوم)
درجه علمی: نشریه علمی (وزارت علوم)
آرشیو
چکیده
یکی از اولویت های مهم وزارت جهاد کشاورزی، تهیه نقشه نوع محصول کشاورزی برای تخمین میزان سطح زیرکشت محصولات استراتژیک و برآورد سالیانه میزان تولید آنهاست. در دهه های اخیر، فناوری سنجش از دور به دلیل تهیه تصاویر و داده های به هنگام با تفکیک پذیری های متنوع مکانی، زمانی و طیفی و با بهره گیری از الگوریتم های یادگیری ماشین بهبودیافته در تخمین میزان سطح زیرکشت محصولات کارایی زیادی را نشان داده است. در پژوهش حاضر با استفاده از سری زمانی تصاویر ماهواره لندست-8 و الگوریتم های یادگیری ماشین پیشرفته یک چهارچوب تهیه نقشه نوع محصول کشاورزی مرودشت استان فارس ارائه شد. الگوریتم های به کار گرفته شده شامل الگوریتم درخت تصمیم، جنگل تصادفی، جنگل دورانی، ماشین بردار پشتیبان و آنالیز انحراف زمانی پویا بود. نتایج نشان داد که روش های آنالیز انحراف زمانی پویا و جنگل تصادفی نسبت به روش های دیگر کارایی بسیار بیشتری (با افزایش دقت کلی به میزان 10% تا 12% بیشتر) در تهیه نقشه نوع محصول کشاورزی منطقه مطالعه شده داشتند. همچنین، در این پژوهش قابلیت باندهای 2 تا 5 ماهواره لندست-8 در شناسایی کارا و مطمئن همه محصولات این منطقه با استفاده از روش های مذکور اثبات شد.Crop Mapping from Landsat-8 Images Time Series Using Machine-Learning Methods (Case Study: Marvdasht in Fars Province)
One of the key priorities of the Ministry of Agriculture Jihad is the mapping of croplands to estimate crop acreage and annual yield. In recent decades, remote sensing technology has proven to be highly effective in estimating the extent of crop cultivation through the use of timely images and synchronized data with diverse spatial, temporal, and spectral resolutions, leveraging advanced machine-learning algorithms. This study presented a framework for crop mapping in Marvdasht, Fars Province, by utilizing time series of Landsat-8 satellite images and advanced machine-learning algorithms. The employed algorithms included Decision Tree (DT), Random Forest (RF), Rotation Forest (RoF), Support Vector Machine (SVM), and Dynamic Time Warping (DTW) analysis. The results indicated that the dynamic time warping and random forest methods outperformed others, achieving significantly higher accuracy (with an overall accuracy improvement of 10-12%) in generating the agricultural land-use map of the study area. Furthermore, this research demonstrated the effectiveness of Bands 2-5 of Landsat-8 satellite in confidently identifying all crops in this region using the mentioned methods. Keywords: Crop Mapping, Crop Acreage Estimation, Landsat-8 Satellite, Machine Learning, Random Forest (RF), Support Vector Machine (SVM), Dynamic Time Warping (DTW), Remote Sensing. IntroductionOne of the key priorities of the Ministry of Agricultural Jihad and the related organizations is to create a cropland map for estimating crop area and annual yield of strategic products. This study aimed to establish a framework based on classification algorithms using time series of Landsat-8 satellite images to estimate the crop acreage of agricultural products. The study area was the city of Marvdasht in Fars Province, which boasted a relatively diverse range of agricultural crop types. In addition to the main satellite bands, the NDVI time series was also utilized for collecting the data in this research. Various classification methods, such as Support Vector Machine (SVM), Decision Tree (DT), Random Forest (RF), and Rotation Forest (RoF), were compared with the Dynamic Time Warping (DTW) method as well. Furthermore, a reliable cropland map of the study area was prepared by using each method. MethodologyIn this study, a rule-based framework utilizing well-known classifiers, such as DT, RF, RoF, SVM, and DTW methods, was proposed by using an 8-day time series of Landsat-8 satellite images and the Digital Elevation Model (DEM) of the region. The proposed framework consisted of two fundamental steps: Firstly, non-agricultural areas, including urban areas, water bodies, and mountainous regions, were masked in the region by using several rules. Subsequently, the agricultural fields were classified. To mask urban areas, the Standard Deviation (SD) of the time series of the NDVI index was employed. For identifying water bodies, a simple thresholding method was used based on the negative NDVI values. Additionally, the DEM of the region was utilized to mask highland areas. Finally, the pixels belonging to agricultural areas were classified by using the five algorithms, i.e., DT, RF, RoF, SVM, and DTW. Research FindingsUsing only the NDVI index, the methods achieved an overall accuracy between 81% and 86% (Kappa coefficient of 0.78-0.83). Notably, the RF and DTW methods exhibited the highest overall accuracy. The resulting SD also indicated the high stability of these two methods compared to the others. Conversely, the RoF method demonstrated the lowest accuracy, while the DT method proved to be less stable than all the other methods. The SVM method, exhibited similar inefficiency and low reliability in identifying summer crops and tomato products as the RoF method although it achieved relatively a good overall accuracy and kappa coefficient. In contrast, the RF and DTW methods were able to distinguish tomato products with relatively good accuracy, but they did not perform well in detecting summer crops. It appeared that regardless of the classification method, relying solely on the NDVI index was insufficient for vegetation identification.Another important observation was that none of these methods achieved the standard accuracy of 85% for identifying products in the study area. Even the RF and DTW methods, which exhibited higher F-score values in most classes compared to the other methods, could not meet this standard accuracy for most classes, including sugar beet, wheat, corn, summer crops, and tomatoes. Therefore, it could be concluded that relying solely on the NDVI index might not be effective for classifying the agricultural croplands of this region.When the time series of the spectral bands were used in conjunction with the NDVI index time series, the overall accuracy of all the methods improved from 88% to 96% (a 6-12% increase in accuracy). The highest increase in accuracy was 12% and 10% for the DTW and RF methods, respectively. Additionally, these two methods exhibited greater stability compared to the other methods. The inclusion of spectral bands even provided additional assistance to the DT, RoF, and SVM methods in identifying certain crops, such as summer crops and tomatoes, which had been challenging to identify in the previous stage. Another noteworthy point was that, with the addition of spectral bands, the standard accuracy of 85% was ultimately achieved for all crop types in the city of Marvdasht by the two methods of DTW and RF. Discussion of Results & ConclusionThis article explored the creation of a cropland map for the agricultural fields of Marvdasht City in Fars Province during the 2014-15 crop year using remote sensing images. An 8-day time series of the spectral bands of 2 to 5 and the NDVI index of Landsat-8 satellite images were employed. This case study encompassed a wide variety of crops, including alfalfa, barley, wheat, corn, sugar beet, tomato, rice, and summer crops. The classification methods utilized in this study comprised Decision Tree (DT), Random Forest (RF), Rotation Forest (RoF), Support Vector Machine (SVM), and Dynamic Time Warping (DTW) algorithms. The implementation of these methods in the mentioned area led to the following conclusions: The DTW and RF methods proved to be more effective than the other methods in identifying agricultural crop types in the study area. Conversely, the DT and RoF methods exhibited subpar performance. The overall low accuracy and inability to identify certain products, such as tomatoes and summer crops, by relying solely on the NDVI index, raised doubts about the efficacy of this method in this region. On the other hand, the DTW method as a novel approach demonstrated excellent efficiency in identifying the agricultural crop types in this region. In summary, this article underscored that, relying solely on the NDVI index for creating a reliable and effective cropland map in Marvdasht was insufficient and it was imperative to incorporate additional features, such as spectral bands, to achieve high accuracy in crop mapping.
تبلیغات
