TITLE:
Influence of Pixel Quality, Land Cover, and Hydroclimatic Cycle on Moderate Resolution Imaging Spectroradiometer Inundation Monitoring Performance in the Pantanal, Brazil
AUTHORS:
Sérgio Wagner Gripp da Silveira, Ibraim Fantin-Cruz, Peter Zeilhofer
KEYWORDS:
MODIS, Classification Performance, Flooding, LSWI, EVI, Large Tropical Wetlands
JOURNAL NAME:
Journal of Geoscience and Environment Protection,
Vol.11 No.2,
February
28,
2023
ABSTRACT: Moderate resolution imaging spectroradiometer
(MODIS) time series (TS) have been widely
applied for flood monitoring in large tropical wetlands. However, little
systematic work is available on the influence of pixel quality, vegetation
cover, and the annual hydroclimatic cycle on classification performance. In
this study, this issue is examined based on a six-year, 250 m resolution
MOD13Q1 TS underpinned by extensive in
situ measurements. The most parsimonious logistic regression model was
obtained for land surface water index (LSWI) and enhanced vegetation index
(EVI). The inclusion of the 500 m MCD12Q1 land cover Type 2 product improves
accuracy. Performance markedly decreases for subsets that include pixels with a
VI quality assurance (QA) level poorer than 0110 and/or a pixel reliability
(PR) of three. When a Savitzky-Golay
filter was used for TS reconstitution, performance is slightly lower than those
obtained in a classification of a VI QA 0001 or PR = 0 level strata; moreover,
these have the advantage of gap-free flood monitoring. The overall accuracy
(OA) of the PR = 0 subset is better for grasslands, and slightly lower for
Savannah, and for woodland and forests. The average OA is highest for the dry
season, intermediate for the rainy/flooded season, and lowest for the
transitional seasons, when the wetland becomes flooded or dries. Comparisons of
internal, k-fold, and external validations indicate that only external
validation enables a realistic assessment of flood-mapping performance. The
complete substitution of PR = 3 pixels by filled-in values is recommended for
operational flood monitoring, and it is concluded that the use of the simplified PR metrics as filtering
criteria for gap filling and smoothing is sufficient for flood
monitoring in the Pantanal. Classification metrics vary more strongly as a
function of the hydrological period than by vegetation cover. MOD13Q1 users
should be aware that OA in forest stands during the transition seasons are, on
average, 25 p.p. lower than the average OAs obtained for the entire series.