Moses Azong Cho

High density biomass estimation for wetland vegetation using WorldView-2 imagery and random forest regression algorithm

a b s t r a c t The saturation problem associated with the use of NDVI for biomass estimation in high canopy density vegetation is a well known phenomenon. Recent field spectroscopy experiments have shown that narrow band vegetation indices computed from the red edge and the NIR shoulder can improve the estimation of biomass in such situations. However, the wide scale unavailability of high spectral resolution satel- lite sensors with red edge bands has not seen the up-scaling of these techniques to spaceborne remote sensing of high density biomass. This paper explored the possibility of estimate biomass in a densely veg- etated wetland area using normalized difference vegetation index (NDVI) computed from WorldView-2 imagery, which contains a red edge band centred at 725 nm. NDVI was calculated from all possible two band combinations of WorldView-2. Subsequently, we utilized the random forest regression algorithm as variable selection and a regression method for predicting wetland biomass. The performance of random forest regression in predicting biomass was then compared against the widely used stepwise multiple linear regression. Predicting biomass on an independent test data set using the random forest algorithm and 3 NDVIs computed from the red edge and NIR bands yielded a root mean square error of prediction (RMSEP) of 0.441 kg/m 2 (12.9% of observed mean biomass) as compared to the stepwise multiple lin- ear regression that produced an RMSEP of 0.5465 kg/m2 (15.9% of observed mean biomass). The results demonstrate the utility of WorldView-2 imagery and random forest regression in estimating and ulti- mately mapping vegetation biomass at high density - a previously challenging task with broad band satellite sensors.

Assessing the utility WorldView-2 imagery for tree species mapping in South African subtropical humid forest and the conservation implications: Dukuduku forest patch as case study

Abstract Indigenous forest biome in South Africa is highly fragmented into patches of various sizes (most patchesxa0 2 ). The utilization of timber and non-timber resources by poor rural communities living around protected forest patches produce subtle changes in the forest canopy which can be hardly detected on a timely manner using traditional field surveys. The aims of this study were to assess: (i) the utility of very high resolution (VHR) remote sensing imagery (WorldView-2, 0.5–2xa0m spatial resolution) for mapping tree species and canopy gaps in one of the protected subtropical coastal forests in South Africa (the Dukuduku forest patch (ca.3200xa0ha) located in the province of KwaZulu-Natal) and (ii) the implications of the map products to forest conservation. Three dominant canopy tree species namely, Albizia adianthifolia , Strychnos spp . and Acacia spp ., and canopy gap types including bushes (grass/shrubby), bare soil and burnt patches were accurately mapped (overall accuracyxa0=xa089.3xa0±xa02.1%) using WorldView-2 image and support vector machine classifier. The maps revealed subtle forest disturbances such as bush encroachment and edge effects resulting from forest fragmentation by roads and a power-line. In two stakeholders’ workshops organised to assess the implications of the map products to conservation, participants generally agreed amongst others implications that the VHR maps provide valuable information that could be used for implementing and monitoring the effects of rehabilitation measures. The use of VHR imagery is recommended for timely inventorying and monitoring of the small and fragile patches of subtropical forests in Southern Africa.

Spectral resampling based on user-defined inter-band correlation filter: C3 and C4 grass species classification

Abstract In this paper, a user-defined inter-band correlation filter function was used to resample hyperspectral data and thereby mitigate the problem of multicollinearity in classification analysis. The proposed resampling technique convolves the spectral dependence information between a chosen band-centre and its shorter and longer wavelength neighbours. Weighting threshold of inter-band correlation (WTC, Pearsons r ) was calculated, whereby r xa0=xa01 at the band-centre. Various WTC ( r xa0=xa00.99, r xa0=xa00.95 and r xa0=xa00.90) were assessed, and bands with coefficients beyond a chosen threshold were assigned r xa0=xa00. The resultant data were used in the random forest analysis to classify in situ C 3 and C 4 grass canopy reflectance. The respective WTC datasets yielded improved classification accuracies (kappaxa0=xa00.82, 0.79 and 0.76) with less correlated wavebands when compared to resampled Hyperion bands (kappaxa0=xa00.76). Overall, the results obtained from this study suggested that resampling of hyperspectral data should account for the spectral dependence information to improve overall classification accuracy as well as reducing the problem of multicollinearity.

Potential of Sentinel-2 spectral configuration to assess rangeland quality

Abstract. Sentinel-2 is intended to improve vegetation assessment at local to global scales. Today, estimation of leaf nitrogen (N) as an indicator of rangeland quality is possible using hyperspectral systems. However, few studies based on commercial imageries have shown a potential of the red-edge band to accurately predict leaf N at the broad landscape scale. We intend to investigate the utility of Sentinel-2 for estimating leaf N concentration in the African savanna. Grass canopy reflectance was measured using the analytical spectral device (ASD) in concert with leaf sample collections for leaf N chemical analysis. ASD reflectance data were resampled to the spectral bands of Sentinel-2 using published spectral response functions. Random forest (RF), partial least square regression (PLSR), and stepwise multiple linear regression (SMLR) were used to predict leaf N using all 13 bands. Using leave-one-out cross validation, the RF model explained 90% of leaf N variation, with a root mean square error of 0.04 (6% of the mean), which is higher than that of PLSR and SMLR. Using RF, spectral bands centered at 705 nm (red edge) and two shortwave infrared bands centered at 2190 and 1610 nm were found to be the most important bands in predicting leaf N.

Potential utility of the spectral red-edge region of SumbandilaSat imagery for assessing indigenous forest structure and health

Indigenous forest degradation is regarded as one of the most important environmental issues facing Sub-Saharan Africa and South Africa in particular. We tested the utility of the unique band settings of the recently launched South African satellite, SumbandilaSat in characterising forest fragmentation in a fragile rural landscape in Dukuduku, northern KwaZulu-Natal. The AISA Eagle hyperspectral image was resampled to the band settings of SumbandilaSat and SPOT 5 (green, red and near infrared bands only) for comparison purposes. Variogram analysis and the red edge shift were used to quantify forest heterogeneity and stress levels, respectively. Results showed that the range values from variograms can quantify differences in spatial heterogeneity across landscapes. The study has also shown that the unique band settings of SumbandilaSat provide additional information for quantifying stress in vegetation as compared to SPOT image data. This is critical in light of the fact that stress levels in vegetation have previously been quantified using hyperspectral sensors, which are more expensive and do not cover large areas as compared to SumbandilaSat satellite. The study moves remote sensing a step closer to operational monitoring of indigenous forests.

Optimizing spectral resolutions for the classification of C3 and C4 grass species, using wavelengths of known absorption features

Abstract. Hyperspectral remote-sensing approaches are suitable for detection of the differences in 3-carbon ( C 3 ) and four carbon ( C 4 ) grass species phenology and composition. However, the application of hyperspectral sensors to vegetation has been hampered by high-dimensionality, spectral redundancy, and multicollinearity problems. In this experiment, resampling of hyperspectral data to wider wavelength intervals, around a few band-centers, sensitive to the biophysical and biochemical properties of C 3 or C 4 grass species is proposed. The approach accounts for an inherent property of vegetation spectral response: the asymmetrical nature of the inter-band correlations between a waveband and its shorter- and longer-wavelength neighbors. It involves constructing a curve of weighting threshold of correlation (Pearson’s r ) between a chosen band-center and its neighbors, as a function of wavelength. In addition, data were resampled to some multispectral sensors—ASTER, GeoEye-1, IKONOS, QuickBird, RapidEye, SPOT 5, and WorldView-2 satellites—for comparative purposes, with the proposed method. The resulting datasets were analyzed, using the random forest algorithm. The proposed resampling method achieved improved classification accuracy ( κ = 0.82 ), compared to the resampled multispectral datasets ( κ = 0.78 , 0.65, 0.62, 0.59, 0.65, 0.62, 0.76, respectively). Overall, results from this study demonstrated that spectral resolutions for C 3 and C 4 grasses can be optimized and controlled for high dimensionality and multicollinearity problems, yet yielding high classification accuracies. The findings also provide a sound basis for programming wavebands for future sensors.

Mapping canopy gaps in an indigenous subtropical coastal forest using high-resolution WorldView-2 data

Invasive species usually colonize canopy gaps in tropical and subtropical forests, which results in a loss of native species. Therefore, an understanding of the location and distribution of canopy gaps will assist in predicting the occurrence of invasive species in such canopy gaps. We tested the utility of WorldView-2 (WV-2) with eight spectral bands at 2 m spatial resolution to delineate forest canopy gaps in a subtropical Dukuduku coastal forest in South Africa. We compared the four conventional visible-near-infrared bands with the eight-band WV-2 image. The eight-band WV-2 image yielded a higher overall accuracy of 86.90% (kappa coefficient = 0.82) than the resampled conventional four-band image that yielded an overall accuracy of 74.64% (kappa coefficient = 0.63) in pixel-based classification. We further compared the vegetation indices that were derived from four conventional bands with those derived from WV-2 bands. The enhanced vegetation index yielded the highest overall accuracy in the category of conventional indices (85.59% at kappa coefficient = 0.79), while the modified plant senescence reflectance index involving the red-edge band showed the highest overall accuracy (93.69%) in the category of indices derived from eight-band WV-2 imagery in object-based classification. Overall, the study shows that the unique high-resolution WV-2 data can improve the delineation of canopy gaps as compared to the conventional multispectral bands.

Predicting C3 and C4 grass nutrient variability using in situ canopy reflectance and partial least squares regression

The use of hyperspectral data to estimate forage nutrient content can be a challenging task, considering the multicollinearity problem, which is often caused by high data dimensionality. We predicted some variability in the concentration of limiting nutrients such as nitrogen (N), crude protein (CP), moisture, and non-digestible fibres that constrain the intake rate of herbivores. In situ hyperspectral reflectance measurements were performed at full canopy cover for C3 and C4 grass species in a montane grassland environment. The recorded spectra were resampled to 13 selected band centres of known absorption and/or reflectance features, WorldView-2 band settings, and to 10 nm-wide bandwidths across the 400–2500 nm optical region. The predictive accuracy of the resultant wavebands was assessed using partial least squares regression (PLSR) and an accompanying variable importance (VIP) projection. The results indicated that prediction accuracies ranging from 66% to 32% of the variance in N, CP, moisture, and fibre concentrations can be achieved using the spectral-only information. The red, red-edge, and shortwave infrared (SWIR) wavelength regions were the most sensitive to all nutrient variables, with higher VIP values. Moreover, the PLSR model constructed based on spectra resampled around the 13 preselected band centres yielded the highest sensitivity to the predicted nutrient variables. The results of this study thus suggest that the use of the spectral resampling technique that uses only a few but strategically selected band centres of known absorption or reflectance features is sufficient for forage nutrient estimation.

Seasonal separation of African savanna components using WorldView-2 imagery : a comparison of pixel- and object-based approaches and selected classification algorithms

Separation of savanna land cover components is challenging due to the high heterogeneity of this landscape and spectral similarity of compositionally different vegetation types. In this study, we tested the usability of very high spatial and spectral resolution WorldView-2 (WV-2) imagery to classify land cover components of African savanna in wet and dry season. We compared the performance of Object-Based Image Analysis (OBIA) and pixel-based approach with several algorithms: k-nearest neighbor (k-NN), maximum likelihood (ML), random forests (RF), classification and regression trees (CART) and support vector machines (SVM). Results showed that classifications of WV-2 imagery produce high accuracy results (>77%) regardless of the applied classification approach. However, OBIA had a significantly higher accuracy for almost every classifier with the highest overall accuracy score of 93%. Amongst tested classifiers, SVM and RF provided highest accuracies. Overall classifications of the wet season image provided better results with 93% for RF. However, considering woody leaf-off conditions, the dry season classification also performed well with overall accuracy of 83% (SVM) and high producer accuracy for the tree cover (91%). Our findings demonstrate the potential of imagery like WorldView-2 with OBIA and advanced supervised machine-learning algorithms in seasonal fine-scale land cover classification of African savanna.

Mapping the occurrence of Chromolaena odorata (L.) in subtropical forest gaps using environmental and remote sensing data.

Globally, subtropical forests are rich in biodiversity. However, the native biodiversity in these forests is threatened by the presence of invasive species such as Chromolaena odorata (L.) King and Robinson, which thrives in forest canopy gaps. Our study explored the utility of WorldView-2 data, an 8-band high resolution (2xa0m) imagery for mapping the probability of C. odorata occurrence (presence/absence) in canopy gaps of a subtropical forest patch, the Dukuduku forest, South Africa. An integrated modelling approach involving the WorldView-2 vegetation indices and ancillary environmental data was also assessed. The results showed a higher performance of the environmental data only model (deviance or D2xa0=xa00.52, pxa0<xa00.05, nxa0=xa077) when compared to modelling with WorldView-2 vegetation indices such as the enhanced vegetation index, simple ratio indices and red edge normalized difference vegetation index (D2xa0=xa00.30, pxa0<xa00.05, nxa0=xa077). The integrated model explained the highest presence/absence variance of C. odorata (D2xa0=xa00.71, i.e. 71xa0%). This model was used to derive a probability map indicating the occurrence of invasive species in forest gaps. A 2xa0×xa02 error matrix table and the receiver operating characteristic curve derived from an independent validation dataset (nxa0=xa038) were used to assess the model accuracy. Approximately 87xa0% of canopy gaps containing C. odorata were correctly predicted at probability threshold of 0.3. The derived probability map of C. odorata occurrence could assist management in prioritizing target areas for eradication of the species.