Browsing by Subject "imaging spectroscopy"
Now showing items 1-2 of 2
-
(2024)Grass biomass has many important and diverse roles for ecosystems functioning, the carbon cycle, rangeland productivity and local livelihoods. Quantifying and understanding grass biomass in dynamic savanna ecosystems during dry season is important for sustainable land management and monitoring grazing pressures, especially amidst climate change. Traditional ground-based methods to assess vegetation are subjective and time consuming, while remote sensing provides efficiency in monitoring grass biomass at large scales. Grass biomass assessments using remote sensing data have been extensively conducted worldwide, but such research in African savannas remains rare. This study aimed to study connections between dry season grass biomass measured in savanna rangelands and airborne hyperspectral imagery data obtained simultaneously in LUMO Conservancy area of South-Eastern Kenya. Two modelling techniques were compared: averaged plot values (n=24) and individual sample values (n=96). Three vegetation indices (RSI, NDSI, RDSI) were computed and Generalised Additive Models (GAM) were applied to portray the relationship between measured grass biomass and VIs. The highest explanatory power for both modelling techniques was found with RSI and NDSI indices with averaged plot level values having the highest performance (D2 = 0.79, RMSE = 40.15 g/m2), with the band combination of B78 and B43 (908 nm / 667 nm). The best performing vegetation index (RSI) was used to predict grass biomass in the study area, which indicated a biomass range of 0 to 2894 g/m2. The study highlights the potential of using hyperspectral imagery to assess grass biomass in the savanna environments. However, challenges and limitations were faced related to the heterogeneous nature of savannas, varying weather conditions affected by rainfall, the temporal limits of the study, and disturbances in spectral information caused by heavily grazed areas, dead material, and preprocessing techniques. It is suggested that future research considers these factors by incorporating a broader set of variables, extending the duration of the study, exploring various preprocessing techniques, increasing the sample size, and employing additional data sources, such as active sensors and hyperspectral satellite imagery, to enhance model performance and improve accuracy.
-
(2020)Soil is the largest actively cycling terrestrial carbon pool, which has been severely distrubed in the last 100-200 years by human actions. To improve the situation, extensive monitoring of soil carbon and new methods for monitoring are required. This study demonstrates the capability of a portable hyperspectral device operating in the visible-near infrared (VIS-NIR) spectrum for soil organic carbon (SOC) prediction. Two multivariate methods, partial least squares regression (PLSR) and for this purpose previously untested lasso regression were used for prediction. 191 soil samples were collected from Taita Hills, Kenya. The samples represent a tropical altitudinal gradient with five land uses: agroforestry, field, forest, shrubland and sisal plantation. The samples were imaged with hyperspectral camera, Specim IQ in laboratory and in field conditions, and the carbon content of the samples was determined with a dry-oxidization analyzer. Three datasets were derived from the images, one containing the mean spectra of the complete imaged samples, one with segmented sub-image spectra and one with segmented sub-image spectra where outlier spectra were removed. Both multivariate methods were tested with all three datasets with good prediction accuracies (PLSR: R2min = 0.85, RMSEmin = 0.78, lasso: R2min=0.85, RMSEmin=0.80), demonstrating the feasibility of both the device and lasso regression as SOC prediction tools. Using the segmented sub-image datasets improved the results with PLSR but had no significant effect on lasso regression prediction results. While good results were gained with laboratory imagery, the field imaging conditions were difficult, and the data performed poorly. Future research should focus on finding solutions to reliably estimate SOC content in situ or with portable laboratory setups to make SOC measurements more widely accessible and agile for e.g. precision agriculture purposes.
Now showing items 1-2 of 2