A New Concept of Soil Line Retrieval from Landsat 8 Images for Estimating Plant Biophysical Parameters

Abstract

Extraction of vegetation information from remotely sensed images has remained a long-term challenge due to the influence of soil background. To reduce this effect, the slope and intercept of the soil line (SL) should be known to calculate SL-related vegetation indices (VIs). These VIs can be used to estimate the biophysical parameters of agricultural crops. However, it is a difficult task to retrieve the SL parameters under the vegetation canopy. A feasible method for retrieving these parameters involves extracting the bottom boundary line in two-dimensional spectral spaces (i.e., red and near-infrared bands). In this study, the slope and intercept of the SL was extracted from Landsat 8 OLI images of a test site in northeastern Germany. Different statistical methods, including the Red-NIR min method, quantile regression method (using a floating tau with the smallest p-value), and a new approach proposed in this paper using a fixed quantile tau known as the diffuse non-interceptance (DIFN) value, were applied to retrieve the SL parameters. The DIFN value describes the amount of light visible below the canopy that reaches the soil surface. Therefore, this value can be used as a threshold for retrieving the bottom soil line. The simulated SLs were compared with actual ones extracted from ground truth data, as recorded by a handheld spectrometer, and were also compared with the SL retrieved from bare soil pixels of the Landsat 8 image collected after harvest. Subsequently, the SL parameters were used to separately estimate the dry biomasses of winter wheat (Triticum aestivum L.), barley (Hordeum vulgare L.), and canola (Brassica napus L.) at the local and field scales using different SL-related vegetation indices. The SL can be retrieved more accurately at the local scale compared with the field scale, and its simulation can be critical in the field due to significant differences from the actual SL. Moreover, the slope and intercept of the simulated SLs found using the floating and fixed quantile tau (slope ≈ 1.1 and intercept ≈ 0.05) show better agreement with the actual SL parameters (slope ≈ 1.2 and intercept ≈ 0.03) in the late growing stages (i.e., end of ripening and senescence stages) of crops. The slope and intercept of the soil line extracted from bare soil pixels of the Landsat 8 OLI data after harvest (slope = 1.3, intercept = 0.03, and R 2 = 0.94) are similar to those of the simulated SL. The correlation coefficient (R 2 ) of the simulated SLs are greater than 0.97 during different growing stage and all of the SL parameters are statistically significant (p < 0.05) at the local scale. The results also imply the need for different vegetation indices to best retrieve the crop biomass depending on the growing stage, but relatively small differences in performances were observed in this study.

A Simple Algorithm for Deriving an NDVI-Based Index Compatible between GEO and LEO Sensors: Capabilities and Limitations in Japan

“…Many factors determined the slope and intercept, including the soil physical and chemical properties, soil moisture, soil mineral composition, roughness, etc. [57]. The equation is called a "soil line-like equation" due to the fact that any soil line is likely to be influenced by vegetation in the 1 km resolution data over our target region.…”

“…This process was also useful for computing boundary lines in two-dimensional reflectance space [60]. The quantile regression has been widely used in soil line estimation [57,60]. The quantile level, τ was set to p 5 (=0.04) in the quantile regression.…”

Estimation of Rangeland Production in the Arid Oriental Region (Morocco) Combining Remote Sensing Vegetation and Rainfall Indices: Challenges and Lessons Learned

“…TSAVI's χ parameter was kept equal to 0.08 [26]. Slope and intercept of the soil line (MSAVI and TSAVI) were inferred from the NIR versus R scatterplot using a quantile regression method [81,82] (Figure S1). For this purpose, we extracted the (R, NIR) couples for all the pixels belonging to the study sites and performed a series of quantile regressions (using the "quantreg" package, version 5.73 [83]) with different tau values, representing the values of the quantiles of the response variable (NIR reflectance) to be estimated by the model given the value of the predictor (R reflectance), and ranging from 1 × 10 −4 to 5 × 10 −3 .…”

Landsat Time-series Images-based Urban Heat Island Analysis: The Effects of Changes in Vegetation and Built-up Land on Land Surface Temperature in Summer in the Hanoi Metropolitan Area, Vietnam

“…Where; ρ is the pixel surface reflectance; ρ e is an average surface reflectance for the pixel and a surrounding region; S is the spherical albedo of the atmosphere; L a * is the radiance backscattered by the atmosphere; A and B are coefficients that depend on atmospheric and geometric conditions but not on the surface. The values of A, B, S and L a * are determined based on MODTRAN4 (Ahmadian et al, 2016). After the water retrieval, the spatially averaged reflectance ρ e is estimated using Equation (4) (Adler-Golden et al, 1998;Adler-Golden et al, 1999):…”