Vegetation Water Content Prediction: Towards More Relevant Explicatory Waveband Variables

Assessing vegetation water content (VWC) from hyperspectral reflectance dataset poses two foremost questions: what specific wavebands of the SWIR offer a good retrieval and what modeling methods have the best predictive ability. In this paper, we explored the application of multivariate statistical techniques such as stepwise multiple linear regression (SMLR) and partial least square regression (PLSR) for vegetation water content prediction using the absorption features. We also examined the recursive partitioning model of the dataset to illustrate relationships of splits among waveband predictors. Previously known wavelength features around 970 nm and 900 nm were leading predictors of water content. In the absence of actual field VWC data, the absorption area feature at 970 nm was computed and used to model other explicatory waveband variables that may boost the prediction. The results of this exploratory waveband-predictor study highlighted other essential locations around 956 nm, 922 nm, 976 nm, 935 nm, and 915 nm. The SMLR disqualified highly correlated bands that augment relatively little in the VWC prediction capability of the model. PLSR presented the 900 nm, 922 nm, and the 970 nm peaks affirming the results of the SMLR. The PLSR is the favored technique with RMSEcv = 0.002 (r2=0.88) for the cross-validation, lower than the SMLR (RMSEcv = 0.023). Recursive partitioning method showed the 956 nm, surprisingly with the highest logworth among predictors. The overall r2 after partitioning, when actual and predicted VWC were plotted against each other was a fair 0.59. This value is comparable with other empirical indices we previously analyzed. Recursive partitioning is a highly adaptive technique and care must be taken in the interpretation of results.