Leveraging satellite observations to reveal ecological drivers of pest densities across landscapes

Landscape ecologists have long suggested that pest abundances increase in simplified, monoculture landscapes. However, tests of this theory often fail to predict pest population sizes in real-world agricultural fields. These failures may arise not only from variations in pest ecology but also from the widespread use of categorical land-use maps that do not adequately characterize habitat availability for pests. We used 1163 field-year observations of Lygus hesperus (Western Tarnished Plant Bug) densities in California cotton fields to determine whether integrating remotely sensed metrics of vegetation productivity and phenology into pest models could improve pest abundance analysis and prediction. Because L. hesperus often overwinters in non-crop vegetation, we predicted that pest abundances would peak on farms surrounded by more non-crop vegetation, especially when the non-crop vegetation is initially productive but then dries down early in the year, causing the pest to disperse into cotton fields. We found that the effect of non-crop habitat on pest densities varied across latitudes, with a positive relationship in the north and a negative one in the south. Aligning with our hypotheses, models predicted that L. hesperus densities were 35 times higher on farms surrounded by high versus low productivity non-crop vegetation (EVI area 350 vs. 50) and 2.8 times higher when dormancy occurred earlier versus later in the year (May 15 vs. June 30). Despite these strong and significant effects, we found that integrating these remote-sensing variables into land-use models only marginally improved pest density predictions in cotton compared to models with categorical land cover metrics alone. Together, our work suggests that the remote sensing variables analyzed here can advance our understanding of pest ecology, but not yet substantively increase the accuracy of pest abundance predictions.

Methods
Our cotton dataset encompassed 1487 field-year replicates of L. hesperus observations across 565 conventionally managed irrigated cotton fields located within 18 ranches (i.e., fields managed by the same organization or grower that may or may not be spatially contiguous). The study site network spanned ~280km of California’s Central Valley, with fields in different ranches separated by an average of 100 km (Interquartile range 31 km). Cotton was usually planted in April (N = 630/872 for which planting date was known). Pesticides were regularly applied to target L. hesperus, most often at peak trap capture (July) and not in the early season studied here (see below). Latitude, longitude, year, and ranch name were available for all fields. Lygus hesperus densities were sampled in Gossypium hirsutum (“upland cotton”) and Gossypium barbadense (“Pima cotton”).

Pest densities were calculated by aggregating 50 swings of a sweep net across the top of the plant canopy. Usually, 6-12 sweep samples were taken for a given field on a given date. Pests were typically surveyed 3-8 times during this early season period (range 1-13) and reflect all motile stages combined. Linear interpolation was used to transform successive samples into mean density estimates by calculating the area under the curve of L. hesperus density by time and dividing by the number of days between sampling intervals. Cotton lint yield was measured and reported once per field year in bales/acre, which was converted to kilograms/hectare for this analysis.

The fractions of crop and non-crop habitat around each focal field were extracted from the National Land Cover Database (NLCD) by quantifying the fraction of 30m2 pixels in each of the two cover classes within three buffer radii around each pest sampling site. Since NLCD data were not available each year, data were matched with the closest year for which data were available (crop years 1997-2002: NLCD 2001, 2003-2005: NLCD 2004, 2006-2007: NLCD 2006, 2008: NLCD 2008). Crop area was defined as either pasture/hay or cultivated crops (NLCD classes 81 and 82). Non-crop vegetation was defined as grasslands (71), shrub/scrub (52), forests (41, 42, 43), or wetlands (90, 95).

We extracted satellite-based climate and vegetation variables within the non-crop habitat. For precipitation in the non-crop habitat, we averaged the total annual precipitation reported from Daymet across all 1 km pixels within both the non-crop habitat and the relevant buffer radius. Daymet data estimate near-surface meteorological conditions where no instrumentation exists using statistically interpolated weather variables.

For information on vegetation productivity throughout the growing season (Enhanced Vegetation Index [EVI] area) and vegetation phenology (dormancy day of the year), we acquired MODIS satellite products (MCD12Q2, Version 6) using the Land Cover Type 2 band. MODIS data are available at a 500 m resolution from 2001-2019; therefore, earlier pest density data (1997-2000) were not analyzed. EVI area reflects the sum of daily estimates of EVI amplitude between green-up and dormancy. The days on which green-up and dormancy are reached were estimated as the days of the year when the EVI amplitude first (green-up) and last (dormancy) crossed 15% of the maximum EVI amplitude. For both EVI area and dormancy of vegetation, values were averaged across all 500 m pixels within the non-crop habitat and the relevant buffer radius. To account for the seasonal nature of precipitation in the California Mediterranean climate, both metrics of productivity (precipitation in non-crop habitat, and EVI area), and phenology (day of year on which dormancy was reached in the non-crop habitat) were estimated using a start date of September 1 in the previous year (i.e., the beginning of the rainy season).

All landscape, precipitation, and satellite observation data were extracted at multiple spatial scales (10 km, 20 km, and 30 km).