Measuring green vegetation cover over agricultural fields: a multi-scale study using smartphones and UAV

Abstract

The fraction of green vegetation cover (fCover) is an important parameter having variety of applications in the field of agriculture and forestry. It refers to the percentage of ground covered with photosynthetically active green parts of a plant. Many methods can be used for the estimation of fCover such as statistical methods relying on vegetation indices, physical methods using the inversion of canopy reflectance models and other. fCover estimates can be retrieved using various instruments and platforms ranging from low cost, consumer grade cameras to professional spectrometers and from unmanned aerial vehicles to satellites. The main objective of this thesis is to study the variation in fCover when estimated at various spatial resolutions, with various instruments and over different extents. Estimates of fCover derived from smartphones and from UAV images were compared. The various estimations were evaluated to better understand the impact of geographic scale on fCover as well as the added value of images with a NIR band. With this analysis, it was possible to estimate the extent with which field measurements agree with the average fCover of a field which can help to design improved fCover field sampling strategies. Next to this, classification and regression methods to estimate fCover were compared. For the classification method, an object-based image analysis approach was adopted which is believed to outperform pixel-based classification techniques in the analysis of high resolution imagery (Blaschke et al., 2014). In the regression method, various vegetation indices and their statistics were used as predictors which allowed the most important color and NIR indices to estimate fCover to be identified. The results showed that minor differences in the resolution of smartphone photos coming from the device models and camera lenses do not have major impact on the fCover estimates. But using too coarse spatial resolution as obtained from UAVs at 65 m altitude can lead to substantial over- or underestimation of fCover. When possible, it is better to use higher resolution photos as obtained from high-end smartphone models because these contribute to the more accurate estimation of fCover. The estimation of fCover based on UAV images also showed that color indices obtained at 12 m altitude were better predictors for fCover than vegetation indices in the near-infrared region obtained at 100 m height. Comparing the two approaches for estimating fCover revealed that both regression and classification algorithms can poorly deal with crops whose leaves are not well defined in respect to the background. Both algorithms had some difficulties estimating fCover when the sun reflection was strong. But the regression algorithm could deal much better with changing illumination when no sun glint was involved. The appropriate number of field samples to be taken depends on the accuracy of fCover sought. Surveying as little as 5% of a field, gave very good fCover estimate that did not differ from the actual fCover with more than 4%.