Kabir Peerbhay

Investigating the Capability of Few Strategically Placed Worldview-2 Multispectral Bands to Discriminate Forest Species in KwaZulu-Natal, South Africa

WorldView-2 multispectral wavebands (8 wavebands; 427-908 nm spectral range; 2 m spatial resolution) were utilized to classify six commercial forest species (Eucalyptus grandis, Eucalyptus nitens, Eucalyptus smithii, Pinus patula, Pinus elliotii and Acacia mearnsii) in South Africa using the partial least squares discriminant analysis (PLS-DA) technique. Results indicate that the WorldView-2 imagery produced an overall accuracy of 85.42% and a kappa statistic value of 0.83, with individual forest species accuracies ranging between 63% and 100%. The variable importance in the projection (VIP) method was then used to identify the most important wavebands that were most effective in discriminating the forest species. Four VIP bands were ranked with thresholds greater than one and produced an overall accuracy of 84.38% and kappa value of 0.81, with individual forest species accuracies between 69% and 100%. More specifically, the VIP bands that were found to be important in the classification were the coastal blue (427 nm), blue (478 nm), green (546 nm) and red (659 nm) and confirmed the relative importance of the visible region of the electromagnetic spectrum in discriminating forest species. Overall, results indicate that multispectral information characterized by greater spatial resolution can successfully discriminate between and within forest species, thus providing an accurate framework for commercial forest species mapping.

Random Forests Unsupervised Classification: The Detection and Mapping of Solanum mauritianum Infestations in Plantation Forestry Using Hyperspectral Data

Detecting and mapping plant invaders using hyperspectral remote sensing is necessary in mitigating the extensive ecologic and economic damage these alien plants induce on our forest ecosystems. Using AISA Eagle image data, this study investigated the capability of two unsupervised classification methods for the detection and mapping of Solanum mauritianum located within commercial forestry ecosystems. The existing random forest (RF) outlier detection method when used in conjunction with Anselins Morans I produced a detection rate (DR) of 89% with a false positive rate (FPR) of 9.26%. In comparison, the newly developed methodology which is based on the decomposition of the RF proximity matrix using principal component analysis (PCA) resulted in a DR of 95% with a lower FPR (6.39%). Overall, this research has demonstrated the potential of utilizing an unsupervised and accurate RF framework for the detection and mapping of alien invasive plants.

Does simultaneous variable selection and dimension reduction improve the classification of Pinus forest species

Abstract Tree species information is important for forest inventory management and supports decisions related to the composition and distribution of forest resources. However, traditional methods of obtaining such information involve time consuming and cost intensive ground-based methods. Hyperspectral data offer an alternative source for obtaining information related to forest inventory. Utilizing Airborne Imaging Spectrometer for Applications Eagle hyperspectral data (393 to 994 nm), this study compares the utility of two partial least squares (PLS)-based methods for the classification of three commercial Pinus tree species. Results indicate that the sparse partial least squares discriminant analysis (SPLS-DA) method performed variable selection and dimension reduction successfully to produce an overall accuracy of 80.21%. In comparison, the PLS-DA method and variable importance in the projection (VIP) selected bands produced an overall accuracy of 71.88%. The most effective bands selected by PLS-DA and VIP coincided within the visible region of the spectrum (393 to 700 nm). However, SPLS-DA selected fewer wavebands within the blue (415 to 483 nm), green (515 to 565 nm), and red regions (674 to 694 nm) to confirm the importance of the visible in discriminating tree species. Overall, this study shows the potential of SPLS-DA to perform simultaneous variable selection and dimension reduction of hyperspectral remotely sensed data resulting in improved classification accuracies.

Comparison of partial least squares and support vector regressions for predicting leaf area index on a tropical grassland using hyperspectral data

Abstract. Leaf area index (LAI) is a key biophysical parameter commonly used to determine vegetation status, productivity, and health in tropical grasslands. Accurate LAI estimates are useful in supporting sustainable rangeland management by providing information related to grassland condition and associated goods and services. The performance of support vector regression (SVR) was compared to partial least square regression (PLSR) on selected optimal hyperspectral bands to detect LAI in heterogeneous grassland. Results show that PLSR performed better than SVR at the beginning and end of summer. At the peak of the growing season (mid-summer), during reflectance saturation, SVR models yielded higher accuracies (R2=0.902 and RMSE=0.371  m2 m−2) than PLSR models (R2=0.886 and RMSE=0.379  m2 m−2). For the combined dataset (all of summer), SVR models were slightly more accurate (R2=0.74 and RMSE=0.578  m2 m−2) than PLSR models (R2=0.732 and RMSE=0.58  m2 m−2). Variable importance on the projection scores show that most of the bands were located in the near-infrared and shortwave regions of the electromagnetic spectrum, thus providing a basis to investigate the potential of sensors on aerial and satellite platforms for large-scale grassland LAI prediction.

Detecting and mapping levels of Gonipterus scutellatus-induced vegetation defoliation and leaf area index using spatially optimized vegetation indices

Abstract Gonipterus scutellatus outbreaks may severely defoliate Eucalyptus plantations growing in South Africa. Therefore, detecting and mapping the severity and extent of G. scutellatus defoliation is essential for the deployment of suppressive measures. In this study, we tested the utility of spatially optimized vegetation indices and an artificial neural network in detecting and mapping G. scutellatus-induced vegetation defoliation, using both visual estimates of percentage defoliation and optical leaf area index (LAI) measures. We tested both field methods to determine which of the two were more superior in detecting vegetation defoliation using optimized vegetation indices. These indices were computed from a WorldView-2 pan-sharpened image, which is characterized with a 0.5-m spatial resolution and eight spectral bands. The indices were resampled to spatial resolutions that best represented levels of G. scutellatus-induced defoliation. The results showed that levels of defoliation, using visual percentage estimates, were detected with an R2 of 0.83 and an RMSE of 1.55 (2.97% of the mean measured defoliation), based on an independent test data-set. Similarly, LAI subjected to defoliation was detected with an R2 of 0.80 and an RMSE of 0.03 (0.06% of the mean measured LAI), based on an independent test data-set. Therefore, the results indicate that the cheaper less-complicated visual percentage estimates of defoliation was the more superior model of the two. A sensitivity analysis revealed that NDRE, MCARI2 and ARI ranked as the top three most influential indices in developing both percentage defoliation and LAI models. Furthermore, we compared the optimized model with a model developed using the original image spatial resolution. The results indicated that the optimized model performed better than the original 0.5-m spatial resolution model. Overall, the study showed that vegetation indices optimized to specific spatial resolutions can effectively detect and map levels of G. scutellatus-induced defoliation and LAI subjected to defoliation.

Multispectral mapping of key grassland nutrients in KwaZulu-Natal, South Africa

Abstract RapidEye multispectral imagery is effective in vegetation assessment. This research intends to investigate the utility of using 5-m-resolution RapidEye-5 imagery combined with a machine learning algorithm to detect and map important forage fibre biochemicals such as neutral detergent fibre (NDF), acid detergent fibre (ADF) and lignin in an African tropical grassland. These fibre biochemicals are significant indicators of the palatability of forage digested by grazing ungulates. Analysis was conducted on 140 grass samples collected at plot level and then correlated and mapped using two competent tree ensemble modelling techniques. Results showed that the random forest (RF) method successfully mapped NDF, ADF and lignin with R2 values ranging between .67 and .74. In comparison, stochastic gradient boosting (SGB) was implemented as an alternative method, which produced very similar R2 values ranging between .65 and 0.72. Overall, the results showed that multispectral remote sensing can detect and map key forage fibre biochemicals in a grassland environment.

The Impact of Simulated Spectral Noise on Random Forest and Oblique Random Forest Classification Performance

Hyperspectral datasets contain spectral noise, the presence of which adversely affects the classifier performance to generalize accurately. Despite machine learning algorithms being regarded as robust classifiers that generalize well under unfavourable noisy conditions, the extent of this is poorly understood. This study aimed to evaluate the influence of simulated spectral noise (10%, 20%, and 30%) on random forest (RF) and oblique random forest (oRF) classification performance using two node-splitting models (ridge regression (RR) and support vector machines (SVM)) to discriminate healthy and low infested water hyacinth plants. Results from this study showed that RF was slightly influenced by simulated noise with classification accuracies decreasing for week one and week two with the addition of 30% noise. In comparison to RF, oRF-RR and oRF-SVM yielded higher test accuracies (oRF-RR: 5.36%–7.15%; oRF-SVM: 3.58%–5.36%) and test kappa coefficients (oRF-RR: 10.72%–14.29%; oRF-SVM: 7.15%–10.72%). Notably, oRF-RR test accuracies and kappa coefficients remained consistent irrespective of simulated noise level for week one and week two while similar results were achieved for week three using oRF-SVM. Overall, this study has demonstrated that oRF-RR can be regarded a robust classification algorithm that is not influenced by noisy spectral conditions.

Can remote sensing detect, monitor and track baboon-damaged Pinus plantations located in South Africa?

Abstract The sustainability of the South African forestry industry is under threat due to the increasing number of biotic stressors in a changing environment. Bark-stripping baboons are impacting the economic potential of the forestry sector, with immediate systems required to quantify plantation damage for real-time adaptive management. This study aims to develop a remotely sensed surveillance system, using the Landsat 8 sensor, to detect and monitor baboon damage over two years. Results show a successful first baboon-damage map regionally, with an overall accuracy of 82.45%. Overall, this research developed an operational framework that supports forest protection using a space-borne satellite platform.

Modelling the susceptibility of pine stands to bark stripping by Chacma baboons (Papio ursinus) in the Mpumalanga Province of South Africa

Abstract Context. Commercial pine (Pinus spp.) plantations in southern Africa have been subjected to bark stripping by Chacma baboons (Papio ursinus) for many decades, resulting in severe financial losses to producers. The drivers of this behaviour are not fully understood and have been partially attributed to resource distribution and availability. Aims. The study sought to develop a spatially explicit ecological-risk model for bark stripping by baboons to understand the environmental factors associated with the presence of damage in the pine plantations of the Mpumalanga province of South Africa. Methods. The model was developed in Random Forests, a machine learning algorithm. Baboon damage information was collected through systematic surveys of forest plantations conducted annually. Environmental predictors included aspects of climate, topography and compartment-specific attributes. The model was applied to the pine plantations of the study area for risk evaluation. Key results. The Random Forests classifier was successful in predicting damage occurrence (F1 score = 0.84, area under curve (AUC) = 0.96). Variable predictors that contributed most to the model classification accuracy were related to pine-stand characteristics, with the age of trees being the most important predictor, followed by species, site index and altitude. Variables pertaining to the environment surrounding a pine stand did not contribute substantially to the model performance. Key conclusions. (1) The study suggests that bark stripping is influenced by compartment attributes; (2) predicted risk of bark stripping is higher in stands above the age of 5 years planted on high-productivity forestry sites, where site index (SI) is above 25; (3) presence of damage is not related to the proximity to natural areas; (4) further studies are required to investigate ecological and behavioural patterns associated with bark stripping. Implications. The model provides a tool for understanding the potential extent of the risk of bark stripping by baboons within this region and it can be applied to other forestry areas in South Africa for risk evaluation. It contributes towards the assessment of natural hazards potentially affecting pine plantations and supports the development of risk-management strategies by forest managers. The model highlights opportunities for cultural interventions that may be tested for damage control.

Remote sensing of key grassland nutrients using hyperspectral techniques in KwaZulu-Natal, South Africa

Abstract. The concentration of forage fiber content is critical in explaining the palatability of forage quality for livestock grazers in tropical grasslands. Traditional methods of determining forage fiber content are usually time consuming, costly, and require specialized laboratory analysis. With the potential of remote sensing technologies, determination of key fiber attributes can be made more accurately. This study aims to determine the effectiveness of known absorption wavelengths for detecting forage fiber biochemicals, neutral detergent fiber, acid detergent fiber, and lignin using hyperspectral data. Hyperspectral reflectance spectral measurements (350 to 2500 nm) of grass were collected and implemented within the random forest (RF) ensemble. Results show successful correlations between the known absorption features and the biochemicals with coefficients of determination (R2) ranging from 0.57 to 0.81 and root mean square errors ranging from 6.97 to 3.03  g/kg. In comparison, using the entire dataset, the study identified additional wavelengths for detecting fiber biochemicals, which contributes to the accurate determination of forage quality in a grassland environment. Overall, the results showed that hyperspectral remote sensing in conjunction with the competent RF ensemble could discriminate each key biochemical evaluated. This study shows the potential to upscale the methodology to a space-borne multispectral platform with similar spectral configurations for an accurate and cost effective mapping analysis of forage quality.