ﮐﺎراﯾﯽ ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﻣﺼﻨﻮﻋﯽ ﺑﻬﯿﻨﻪ در ﻣﺪل ﺳﺎزي ﮐﺮﺑﻦ آﻟﯽ ﺧﺎك ﻣﺒﺘﻨﯽ ﺑﺮ داده ﻫﺎي ﻣﯿﺪاﻧﯽ و ﺗﺼﺎوﯾﺮ 2-Sentinel در ارﺳﺒﺎران

Efficiency of optimized artificial neural network in soil organic carbon modeling based on in-situ measurements and Sentinel-2 images in Arasbaran

ﺳﺎﺑﻘﻪ و ﻫﺪف: ﺧﺎك ﺑﺰرگ ﺗﺮﯾﻦ ﻣﻨﺒﻊ ذﺧﯿﺮه ﮐﺮﺑﻦ ﻣﻮﺟﻮد در ﺑﻮم ﺳﺎزﮔﺎن ﻫﺎي زﻣﯿﻨﯽ ﻫﺴﺘﻨﺪ ﮐﻪ ﺑﯿﺶ ﺗﺮﯾﻦ ﺳﻬﻢ از ﮐﻞ ذﺧﺎﯾﺮ ﺟﻬﺎﻧﯽ ﮐﺮﺑﻦ زﻣﯿﻨﯽ را در ﺧﻮد ﺟﺎي دادﻧﺪ. ﻧﻘﺸﻪ ﺑﺮداري دﻗﯿﻖ اﻃﻼﻋﺎت ﺗﻮزﯾﻊ ﻣﮑﺎﻧﯽ ذﺧﯿﺮه ﮐﺮﺑﻦ آﻟﯽ ﺧﺎك )SOC( ﯾﮏ ﭘﯿﺶ ﻧﯿﺎز ﮐﻠﯿﺪي ﺟﻬﺖ ﻣﺪﯾﺮﯾﺖ ﻣﻨﺎﺑﻊ ﺧﺎك و ﺣﻔﺎﻇﺖ از ﻣﺤﯿﻂ زﯾﺴﺖ اﺳﺖ. ﺗﻮﺳﻌﻪ ﺳﺮﯾﻊ ﻋﻠﻢ ﺳﻨﺠﺶ از دور و اﺳﺘﻔﺎده از ﺗﺼﺎوﯾﺮ ﻣﺎﻫﻮاره اي اﻣﮑﺎن ﻧﻈﺎرت ﺑﺮ ذﺧﯿﺮه SOC در ﻣﻘﯿﺎس ﺑﺰرگ را ﻓﺮاﻫﻢ ﻣﯽ ﮐﻨﺪ. اﻣﮑﺎن ﺑﺮآورد SOC ﯾﮑﯽ از ﻣﻮﺿﻮﻋﺎت ﭘﯿﺶ روي ﭘﮋوﻫﺸﮕﺮان ﺑﻮده اﺳﺖ ﮐﻪ در ﺑﺮﺧﯽ ﻣﻮارد از ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﻣﺼﻨﻮﻋﯽ ﺑﺮاي اﯾﻦ ﻣﻮﺿﻮع اﺳﺘﻔﺎده ﺷﺪه اﺳﺖ ﻫﺮ ﭼﻨﺪ ﺗﻌﯿﯿﻦ ﻣﻘﺎدﯾﺮ ﺑﻬﯿﻨﻪ ﻣﺆﻟﻔﻪ ﻫﺎي ﻣﺆﺛﺮ در آن دﺷﻮار اﺳﺖ. در ﺑﺮﺧﯽ ﻣﻄﺎﻟﻌﺎت از اﻟﮕﻮرﯾﺘﻢ ژﻧﺘﯿﮏ ﺑﺮاي ﺑﻬﯿﻨﻪ ﺳﺎزي وزن ﻫﺎي اوﻟﯿﻪ ﺷﺒﮑﻪ ﻋﺼﺒﯽ و ﺑﻬﺒﻮد ﭘﯿﺶ ﺑﯿﻨﯽ ﻣﺘﻐﯿﺮﻫﺎي ﺧﺮوﺟﯽ اﺳﺘﻔﺎده ﺷﺪه اﺳﺖ. اﮔﺮﭼﻪ ﮐﺎراﯾﯽ اﯾﻦ روش در ﺑﺮآورد SOC ﺑﺎ داده ﻫﺎي ﺳﻨﺠﺶ از دور ﮐﻢ ﺗﺮ ﻣﻮرد ﺑﺮرﺳﯽ ﻗﺮار ﮔﺮﻓﺘﻪ اﺳﺖ. در اﯾﻦ ﭘﮋوﻫﺶ اﺛﺮ اﻟﮕﻮرﯾﺘﻢ ژﻧﺘﯿﮏ ﺑﺮ ﺑﻬﺒﻮد ﻋﻤﻠﮑﺮد ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﻣﺼﻨﻮﻋﯽ در ﭘﯿﺶ ﺑﯿﻨﯽ SOC ﺑﺎ اﺳﺘﻔﺎده از ﺗﺼﺎوﯾﺮ ﻣﺎﻫﻮاره 2-Sentinel در ﻧﺎﺣﯿﻪ روﯾﺸﯽ ارﺳﺒﺎران ﻣﻮرد ﺑﺮرﺳﯽ ﻗﺮار ﮔﺮﻓﺘﻪ اﺳﺖ. ﻣﻮاد و روش ﻫﺎ: ﺑﺮاي اﯾﻦ ﻣﻨﻈﻮر ﻧﻤﻮﻧﻪ ﺑﺮداري از ﺧﺎك ﺑﺎ اﺳﺘﻔﺎده از روش ﻧﻤﻮﻧﻪ ﺑﺮداري ﺗﺼﺎدﻓﯽ ﻃﺒﻘﻪ ﺑﻨﺪي ﺷﺪه ﺑﺮ اﺳﺎس اﻧﻮاع ﮐﺎرﺑﺮي اراﺿﯽ در 46 ﻧﻘﻄﻪ و در ﻋﻤﻖ ﺻﻔﺮ ﺗﺎ 15 ﺳﺎﻧﺘﯽ ﻣﺘﺮ اﻧﺠﺎم ﺷﺪ. SOC ﺑﺎ اﺳﺘﻔﺎده از روش واﻟﮑﻠﯽ ﺑﻼك اﻧﺪازه ﮔﯿﺮي ﺷﺪ. ﺑﺮاي ﺑﺮازش ﻣﺪل ﺑﯿﻦ ﮐﺮﺑﻦ آﻟﯽ اﻧﺪازه ﮔﯿﺮي ﺷﺪه در آزﻣﺎﯾﺸﮕﺎه ﺑﺎ 9 ﺷﺎﺧﺺ ﻃﯿﻔﯽ و 3 ﺑﺎﻧﺪ ﺗﺼﻮﯾﺮ ﻣﺎﻫﻮاره اي ﮐﻪ ﺑﻪ ﻃﻮر ﻣﺴﺘﻘﯿﻢ وارد ﻣﺪل ﺳﺎزي ﺷﺪﻧﺪ، از دو روش ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﻣﺼﻨﻮﻋﯽ و ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﻣﺼﻨﻮﻋﯽ ﺑﻬﯿﻨﻪ ﺷﺪه ﺑﺎ اﻟﮕﻮرﯾﺘﻢ ژﻧﺘﯿﮏ اﺳﺘﻔﺎده ﮔﺮدﯾﺪ. ﺑﺮاي ارزﯾﺎﺑﯽ ﮐﺎراﯾﯽ ﻣﺪل ﻫﺎ از روش اﻋﺘﺒﺎرﺳﻨﺠﯽ ﻣﺘﻘﺎﺑﻞ )Cross Validation( اﺳﺘﻔﺎده ﺷﺪ. در ﻧﻬﺎﯾﺖ ﻣﺪل ﻫﺎي ﺑﻪ دﺳﺖ آﻣﺪه ﺑﺎ ﺷﺎﺧﺺ ﻫﺎي آﻣﺎري ﺟﺬر ﻣﯿﺎﻧﮕﯿﻦ ﻣﺮﺑﻌﺎت ﺧﻄﺎ )RMSE(، ﻧﺴﺒﺖ ﻋﻤﻠﮑﺮد ﺑﻪ اﻧﺤﺮاف )RPD(، ﺿﺮﯾﺐ ﻫﻤﺒﺴﺘﮕﯽ اﺳﭙﯿﺮﻣﻦ )r(، 2 ﺿﺮﯾﺐ ﺗﺒﯿﯿﻦ )R ( و ﻫﻢ ﭼﻨﯿﻦ آزﻣﻮن ﺗﯽ ﺟﻔﺘﯽ ﻣﻮرد ارزﯾﺎﺑﯽ ﻗﺮار ﮔﺮﻓﺘﻨﺪ. ﯾﺎﻓﺘﻪ ﻫﺎ: ﻧﺘﺎﯾﺞ ﻧﺸﺎن داد SOC ﺑﺮآوردي ﺑﺎ ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﻣﺼﻨﻮﻋﯽ ﺑﻬﯿﻨﻪ ﺷﺪه ﺑﺎ اﻟﮕﻮرﯾﺘﻢ ژﻧﺘﯿﮏ )1/07=RMSE درﺻﺪ، 1/89=RPD درﺻﺪ، R2=0/76( دﻗﺖ ﺑﯿﺶ ﺗﺮي ﻧﺴﺒﺖ ﺑﻪ ﻧﺘﺎﯾﺞ ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﻣﺼﻨﻮﻋﯽ )1/51=RMSE درﺻﺪ، 1/34=RPD درﺻﺪ، R2=0/58( داﺷﺖ. ﻫﻢ ﭼﻨﯿﻦ ﺑﻬﺒﻮد ﺿﺮﯾﺐ ﻫﻤﺒﺴﺘﮕﯽ اﺳﭙﯿﺮﻣﻦ ﺑﺮاي SOC واﻗﻌﯽ و ﺑﺮآورد ﺷﺪه ﺑﺎ ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﻣﺼﻨﻮﻋﯽ ﺑﻬﯿﻨﻪ ﺷﺪه )0/87=r( و SOC ﺑﺮآورد ﺷﺪه ﺑﺎ ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﻣﺼﻨﻮﻋﯽ )0/76=r( ﻣﺸﺎﻫﺪه ﮔﺮدﯾﺪ. SOC واﻗﻌﯽ ﺑﺎ SOC ﺑﺮآوردي ﺑﺎ ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﻣﺼﻨﻮﻋﯽ ﺑﻬﯿﻨﻪ اﺧﺘﻼف ﻣﻌﻨﯽ داري ﻧﺪاﺷﺖ )0/21=p-value( وﻟﯽ ﺑﺎ SOC ﺑﺮآوردي ﺑﺎ ﺷﺒﮑﻪ ﻋﺼﺒﯽ ﻣﺼﻨﻮﻋﯽ ﻣﺘﻔﺎوت ﺑﻮد )0/02=p-value(. ﻋﻼوه ﺑﺮ اﯾﻦ، ﻧﺘﺎﯾﺞ ﻧﺸﺎن داد ﮐﻪ ﺷﺎﺧﺺ TSAVI ﺑﯿﺶ ﺗﺮﯾﻦ ﺿﺮﯾﺐ ﻫﻤﺒﺴﺘﮕﯽ اﺳﭙﯿﺮﻣﻦ )0/565( و ﺷﺎﺧﺺ BI2 ﮐﻢ ﺗﺮﯾﻦ ﺿﺮﯾﺐ ﻫﻤﺒﺴﺘﮕﯽ اﺳﭙﯿﺮﻣﻦ )0/196( را ﺑﺎ ﮐﺮﺑﻦ آﻟﯽ ﺧﺎك دارﻧﺪ. ﻧﺘﯿﺠﻪ ﮔﯿﺮي: به طوركلي، نتيجه گيري شد كه استفاده از الگوريتم ژنتيك در انتخاب مولفه هاي بهينه شبكه عصبي مصنوعي منجر به بهبود عملكرد اين روش مدل سازي در برآورد نقطه اي SOC با استفاده از تصاوير ماهواره Sentinel-2 در منطقه موردمطالعه شده است. همچنين با توجه به نتايج بدست آمده، كارايي تصاوير ماهواره Sentinel-2 در برآورد SOC در منطقه مورد مطالعه تاييد شد.

Background and Objectives: Soils are the largest carbon pool in terrestrial ecosystems, which account for the greatest amount of the global total terrestrial carbon stocks. Accurate mapping of Soil Organic Carbon (SOC) spatial distribution is a key assumption for soil resource management and environmental protection. The rapid development of remote sensing and application of satellite images provide an excellent opportunity to monitor large-scale SOC storage. Estimating SOC is one of the research topics that artificial neural networks are applied for this purpose in some studies, although parameter optimization is difficult. In previous studies, genetic algorithms have been used to optimize the artificial neural network initial weights and improve the prediction of the output variables. However, the effectiveness of this method in estimating the SOC by remote sensing has been less studied. In this study, the effect of genetic algorithm on artificial neural network training to predict SOC on Sentinel-2 satellite images in Arasbaran vegetation zone was investigated. Materials and methods: For this purpose, soil sampling was performed using stratified sampling method at 46 points at a depth of 0 to 15 cm. SOC was measured by Walky-Black titration method. In order to fit the model between the measured organic carbon in the laboratory, 9 spectral indices and three bands of satellite image, and two methods were used namely, artificial neural network and artificial neural network optimized by genetic algorithm. Cross validation was used to evaluate the models efficiently. Finally, the precision of the obtained models was evaluated with statistical indices of Root Mean Square Error (RMSE), Ratio of Performance to Deviation (RPD), Spearman's correlation coefficient (r), coefficient of determination (R2), and paired sample t-test. Results: The results showed that the precision of SOC estimated by artificial neural network optimized by genetic algorithm (RMSE = 1.07%, RPD = 1.89%, R2 = 0.76) was higher than artificial neural network results (RMSE =1.51%, RPD = 1.34%, R2 = 0.58). Also Spearman correlation coefficient for SOC estimated with optimized artificial neural network (r = 0.87) was higher compared to estimated SOC with artificial neural network (r = 0.76). Observed SOC was not significantly different from SOC estimated by optimized artificial neural network (p-value=0.21) while it was different from estimated SOC by artificial neural network (p-value=0.02). In addition, the results showed that TSAVI index had the highest Spearman correlation coefficient (0.565), and BI2 index had the lowest Spearman correlation coefficient (0.196) with soil organic carbon. Conclusion: Generally, it was concluded that the use of genetic algorithm in the selection of artificial neural network parameters improved the performance of this modeling method in estimating soil organic carbon on Sentinel-2 satellite images in the study area. Also the performance of Sentinel-2 satellite images in estimating soil organic carbon in the study area was validated