Imputation of HLS image time series for low-latency agricultural applications
Abstract
Growers rely on frequently updated data layers in order to be proactive and informed with their daily decision-making, especially during the growing season. In addition to weather, satellite-derived vegetation indices such as NDVI can be a very useful source of information about crop health and growing stages. The Harmonized Landsat Sentinel-2 (HLS) product provides observations at spatial resolutions high enough to resolve sub-field processes, but the latency remains on a weekly time scale given the overpass times of Landsat 8 and Sentinel-2 sensors, as well as cloud and snow coverage. On the other hand, MODIS-based imagery can be available daily, but the spatial resolution of 250+ m is too coarse to work on a field scale. To bridge this gap, we developed statistical solutions based on the synthesis of MODIS and HLS data to generate vegetation indices with the lowest latency and highest spatial resolution possible for supporting analysis and decision making for agriculture in near-real time . Our approach was to isolate the statistical problem from the interpretation problem by focusing on the creation of a "best possible" but sensor-specific data set, upstream of analysis and modeling solutions. Several potential approaches were explored: (1) simple interpolation, predicting the value at a specific time, based on the other values in the time series; (2) statistical modeling, utilizing covariates available at a specific time; and (3) spatial unmixing, using mean annual time series for each crop type and co-located MODIS pixel. Combinations of these approaches can be used, e.g. the prediction from interpolation, then transformation of MODIS indices based on the co-located HLS time series and/or the mean HLS crop signal can be used in a statistical approach. One guiding principle was the desire to calculate goodness of fit and out of sample error statistics to quantify the added value of each new approach. Thus, we implemented a cross validation strategy, evaluating predicted values at a selected set of times and places for which HLS data are held out.
- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2019
- Bibcode:
- 2019AGUFM.B14A..03M
- Keywords:
-
- 0410 Biodiversity;
- BIOGEOSCIENCES;
- 0480 Remote sensing;
- BIOGEOSCIENCES;
- 1910 Data assimilation;
- integration and fusion;
- INFORMATICS