نوع مقاله : مقاله پژوهشی
نویسندگان
1 گروه علوم و مهندسی خاک، دانشکده کشاورزی، دانشگاه کردستان، سنندج، ایران
2 دانشیار گروه علوم و مهندسی خاک دانشگاه کردستان
3 گروه مدیریت مناطق خشک و بیابانی، دانشکده کشاورزی و منابع طبیعی، دانشگاه اردکان، اردکان، ایران.
چکیده
کلیدواژه‌ها
موضوعات
عنوان مقاله [English]
نویسندگان [English]
The rapid, accurate, and low-cost determination of soil properties has particularly important for land planning and management. The objective of this study was to evaluate the Vis-NIR spectral reflectance of soils, as a rapid, cost-effective, and non-destructive technique, for estimating some soil properties [sand, silt, clay, pH, and calcium carbonate equivalent (CCE)] by partial least-square regression (PLSR) and bagging-PLSR methods. For this purpose, a total of 220 composite soil samples were collected from 0-20 cm depth in Ghorveh Plain, Kurdistan province, in September 2019. The selected soil properties were measured by standard laboratory methods. The proximal spectral reflectance of soil samples was also measured within the 350-2500 nm range (Vis-NIR) using a handheld spectroradiometer. Different pre-processing methods were assessed after recording the spectra. The results indicated that the R2 values for the PLSR method ranged from 0.58 to 0.76, while the bagging-PLSR produced R2 values between 0.59 and 0.74. The RMSE values obtained for sand, silt, clay, CCE, and pH were 17.43, 7.65, 7.83, 7.94, and 0.66, respectively for the PLSR, and 16.66, 7.63, 8.13, 7.71, and 0.45 for the bagging-PLSR. Based on the ratio of prediction to deviation (RPD) values, the bagging-PLSR model achieved the best performance in predicting sand and CCE. However, for clay and pH prediction, the PLSR model was the most accurate. Both the PLSR and bagging-PLSR models yielded identical predictions for silt content, with an RPD value of 1.53. Overall, the results showed that PLSR and bagging-PLSR models have acceptable accuracy for estimating the proposed properties of the soils.
کلیدواژه‌ها [English]
Evaluation of PLSR and bagging-PLSR methods in estimating soil texture, pH, and calcium carbonate using spectral data
EXTENDED ABSTRACT
The rapid, accurate, and low-cost determination of soil properties has particularly important for land planning and management. However, conventional soil sampling and reliable measurement of soil properties, especially on a large geographic scale, can be a laborious task, time-consuming, expensive, and require quantities of harmful chemicals substance for performing experiments. It is, perhaps, for these reasons that the proximal and remote sensing techniques are being considered as possible alternatives to enhance, complement or substitute traditional soil analysis methods. During the last few decades, the use of visible (Vis) and near-infrared (NIR) diffuse reflectance spectroscopy, as a proximal sensing technique, has attracted tremendous attention for assessing soil properties. Therefore, the objective of this study was to evaluate the Vis-NIR spectral reflectance of soils, as a rapid, cost-effective and non-destructive technique, for estimating some soil properties by PLSR and bagging-PLSR methods
A total of 220 composite soil samples were collected from the 0–20 cm depth in Ghorveh Plain, Kurdistan province, in September 2019. These soil samples were transported to the laboratory, air dried, grounded, and then sieved to a size fraction of smaller than 2-mm. The selected soil properties including sand, silt, clay, pH, and calcium carbonate equivalent (CCE) were measured by standard laboratory methods. In addition, the proximal spectral reflectance of soil samples was also measured within the 350-2500 nm range (Vis-NIR) using a handheld spectroradiometer. To minimize the impact of random noise and improve calibration models, different pre-processing methods were assessed after recording the spectra.
The study found that the first derivative of the Savitzky-Golay smoothing filter was the most effective pre-processing technique for calibrating the PLS regression. The optimal number of factors for predicting sand, silt, clay, pH, and CCE were 14, 7, 23, 10, and 18, respectively, using the PLSR method. The R2 values for the PLSR method ranged from 0.58 to 0.76, while the bagging-PLSR produced R2 values between 0.59 and 0.74. The RMSE values obtained for sand, silt, clay, CCE, and pH were 17.43, 7.65, 7.83, 7.94, and 0.66, respectively for the PLSR, and 16.66, 7.63, 8.13, 7.71, and 0.45 for the bagging-PLSR. Based on the ratio of prediction to deviation (RPD) values, the bagging-PLSR model achieved the best performance in predicting sand and CCE. However, for clay and pH prediction, the PLSR model was the most accurate. Both the PLSR and bagging-PLSR models yielded identical predictions for silt content, with an RPD value of 1.53.
It is concluded that the bagging-PLSR outperformed PLSR in predicting sand content and CCE. However, PLSR was more effective for predicting clay content and pH. Both models produced similar results for silt content. In all, the accuracy levels of both PLSR and bagging-PLSR were high for clay content and moderate to good for sand and silt contents, pH, and CCE. These findings suggest that the Vis-NIR spectroscopy, as a complement or replacement approach to laboratory conventional methods, can be used for rapid and cost-efficient assessment of soil properties.
)