ارزیابی چند شاخص طیفی برای برآورد عملکرد کلزا با استفاده از تصاویر سنجنده سنتینل-2

سنجش از دور و به‌کارگیری تصاویر ماهواره‌ها به‌علت سرعت کار و گستردگی سطح پوشش بسیار مورد توجه قرار گرفته است. کلزا به‌دلیل گل‌های زرد آن دارای رنگ پوشش گیاهی متفاوتی با سایر محصولات است و تحقیقات کمی در زمینه ارزیابی شاخص‌های طیفی به‌منظور پیش‌بینی عملکرد آن انجام گردیده است. در سال زراعی 96-95 با هدف پیش‌بینی عملکرد کلزا ده شاخص طیفی سنجنده سنتینل2، مورد ارزیابی قرار گرفت. این تحقیق به شکل پیکسل‌مبنا در سه مزرعه انجام شد و محدوده شبکه‌ای پیکسل‌های مزارع با کمک سیستم موقعیت‌یابی جهانی سینماتیک زمان واقعی (rtkgps) تعیین گردید. در این تحقیق مدل‌های رگرسیونی خطی ساده و چند متغیره و نیز شبکه عصبی به‌کار رفت. نتایج نشان داد براساس مدل رگرسیون خطی ساده، بین مراحل مختلف رشد، بیشترین ضریب تبیین (r^2) در هر یک از شاخص‌های گیاهی در یکی از دو مرحله اوج گل‌دهی و رسیدگی سبز رخ می‌دهد. براساس این مدل، در مرحله اوج گل‌دهی، شاخص تفاضل نرمال شده زردی (ndyi) با 73 درصد بیشترین ضریب تبیین را نسبت به سایر شاخص‌ها احراز کرد. با به‌کارگیری مدل رگرسیون خطی چند متغیره گام به گام با ورودی چهار باند، سه باند مرئی و باند مادون قرمز نزدیک، بهترین مدل در مرحله اوج گل‌دهی با ضریب تبیین 76 درصد و اعتبارسنجی 73 درصد با ریشه میانگین مربعات خطا (rmse) به‌میزان 0.641 به‌دست آمد. همچنین با استفاده از مدل شبکه عصبی و ورود چهار باند مذکور نیز بهترین مدل در مرحله اوج گل‌دهی با ضریب تبیین 92 درصد (آموزش) و اعتبارسنجی (آزمون) 77 درصد با rmse به‌میزان 0.612 احراز شد.

Evaluation of Several Spectral Indices for Estimation of Canola Yield using Sentinel-2 Sensor Images

IntroductionRemote sensing and using satellite images have been widely considered due to the high speed of measurement and great area of coverage. Canola is a source of edible oil and its cultivation in Iran and the world is developing. Comparing with other crops, canola, because of its yellow flowers, has a different canopy color, and only a few researches have been carried out in order to assess the spectral indices for prediction of its yield. Therefore, the main objective of this research is to evaluate some spectral vegetation indices to estimate the yield of canola in different growth stages.Materials and MethodsThe study was performed in 20162017 in Karaj, Iran. Three canola farms were chosen for the evaluation of the relationship between yield and some vegetation indices derived from the Sentinel2 sensor. The sensor data were processed in five stages: before flowering, early flowering, peak of flowering, green and dry maturity, and the vegetation indices were extracted for each of them. This research was pixelbased and the pixels network of each studied farm was determined by RTKGPS. During harvesting time, for measurement of grain yield, five samples, four from the corners and one from the center of the pixel, were taken and their average was considered as the representative amount of the pixel. Totally, 112 pixels from three studied farms were used to calibrate the predictive models. By using Simple Linear Regression (SLR) models, ten new and conventional vegetation indices were assessed. Also, Multivariate Linear Regression (MLR) models and Artificial Neural Net (ANN) models with four bands, three visible bands and NIR band, as inputs, were used to predict the canola yield. In order to validate the SLR and MLR models, the KFold method of crossvalidation was used, and for the validation of ANN models, 15% of data were used; 70% for the train, 15% for validation, and 15% for the test.Results and DiscussionThe results showed that, on the basis of SLR models, among the growth stages, the highest coefficient of determination (R2) in each of the vegetation indices belonged to one of the two stages: the peak of flowering and green maturity. According to SLR models, among the vegetation indices in different stages, the NDYI in the peak of the flowering stage had the highest correlation with yield (R2 = 73%). Also, the RVI with 29%, BNDVI with 52%, NDVI with 56%, and GNDVI with 35% had the highest R2 in the before flowering, early flowering, peak of flowering, green and dry maturity stages, respectively. MLR models resulted to the best yield predictive model at the peak of flowering stage (R2 = 76% for the calibration and R2 = 73% and RMSE = 0.641 for the validation). For ANN models, the strongest model achieved at peak of flowering stage (R2 = 92% for the calibration (train) and R2 = 77% and RMSE = 0.612 for the validation (test)). It seems that the results are affected by yellow flowers of canola, and absorption of blue light by their petals. Therefore, in the peak of the flowering stage, the reflection of the blue light is more likely to belong to green leaves and stems. Therefore, any index such as NDYI, which the blue reflection is subtracted in its equation, represents better the number of flowers, and since the density of flowers is directly related to the yield, the yield will be predicted with more precision. ConclusionsThe results of the analysis of the indices by SLR models showed that the correlation of each of the vegetation indices with the canola yield, in different stages of growth, has a considerable difference. Based on this model, the highest R2 in each of these indices happened in the peak of flowering or green maturity stage, and among these indices in different stages, the NDYI in the peak of the flowering stage had the highest R2. Finally, in both of the MLR and ANN models, with four bands, three visible bands and nearinfrared band, as inputs, the best yield predictive model resulted in the peak of the flowering stage.