Calibration and Validation of Remotely Sensed Ground Cover Maps

Abstract

Calibration and validation is essential in the development of remotely sensed fractional ground cover maps to ensure their reliability and provide users with confidence. Field measurements of fractional cover (FC) are typically collected through surveys where participants have the potential to introduce biases as they categorise ground cover. Environmental factors also have potential to influence the reliability of image-derived products. FC maps have been found to provide poor estimates of cover in arid regions of Australia, and it has been suggested that this may be due to soil colour. Further investigation is required to determine if soil colour influences satellite-derived FC products and there is scope to explore other methods of collecting field measurements in order to reduce errors. The aim of this thesis was to investigate methods of improving fractional ground cover mapping in Australia. The objectives were to (1) trial hyperspectral ground cover sampling in arid Australia by determining how spectral surveys and traditional sampling compare at the same scale and to compare these field methods to satellite-derived FC products, (2) examine observer consistency when classifying vegetation as photosynthetic or non-photosynthetic and to examine how spectral classification of vegetation compares to observer results, and (3) determine if the Australian MODIS FC product is influenced by soil colour. For objective one a sampling design suitable for the evaluation of coarse resolution imagery was developed. Sites were sampled collecting hyperspectral reflectance measurements and step-point observations of ground cover that were later compared to Australian MODIS and Landsat FC products. The results showed a strong relationship between the field sampling methods, that the Landsat FC product was strongly correlated to non-photosynthetic vegetation and soil and the MODIS product was strongly correlated to photosynthetic vegetation. This study demonstrated the hyperspectral field sampling’s improved objectivity, ease of use, and ability to produce results comparable to traditional transect measures. Objective two examined photographs and reflectance measurements of vegetation transitioning from 100% photosynthetic to 100% non-photosynthetic. Observers classified leaves as either photosynthetic or non-photosynthetic (as required in field fractional cover methods), while spectral unmixing was used to decompose the reflectance measurements into photosynthetic and non-photosynthetic proportions. At the extremes (≤ 25 % or ≥ 75 %) photosynthetic observers tended to agree and assigned the leaf to the correct category. However, for leaves in transition (> 25 % or < 75 % photosynthetic) decisions differed more widely and classifications showed little agreement with the spectral proportions of photosynthetic and non-photosynthetic vegetation. This study increased our understanding of the limitations of field data collected using traditional observation methods, of observer variation, and of when observer data may become unreliable. Objective three compared MODIS and TERN AusPlot field estimates of FC at 250 sites across Australia and examined the effect of soil colour (represented by Munsell hue) on the FC values. Overall, there was a significant difference between all 250 sites based on hue suggesting that soil colour has a significant effect on the MODIS product. This evaluation provided insights into the association of specific soil colours with bias in MODIS ground cover fractions and highlighted hues that are associated with under- or overestimation of MODIS FC. Future research may utilise this information to help develop methods of minimising the effects of soil colour in future FC products. This thesis has contributed toward efforts to improve the collection of ground cover measurements for the validation of remotely sensed products, using spectral transect surveys as an alternative to traditional surveys, for photosynthetic activity, provided insight into observer classification consistency and determined how observer-based classification and hyperspectral unmixing compare, and contributed to our understanding of the effects of soil colour on the MODIS FC product. This knowledge will allow informed consumption of the current MODIS FC product, and assist future efforts to calibrate and validate FC products ensuring end-users have reliable and consistent ground cover data for research and decision making.