Large-scale remote sensing based methods for glyphosate usage detection: A case study in the Netherlands

Abstract

Glyphosate is a widely used herbicide, and given its potential threats to health and the environment, it is necessary to monitor its actual use. Currently there is no publicly available remote sensing method for detecting glyphosate use on a large scale. This study developed two glyphosate detection methods based on Sentinel-2 data and compared them in a case study over the entire Netherlands. The NDVI-method detects glyphosate use based on parcel-level NDVI variation, employing multiple constraints. The Color-method detects glyphosate use by analyzing parcel-level changes in spectral values, combining a random forest classifier with multi-temporal constraints. Training and validation data consisted of citizen-science observations from waarneming.nl platform and random parcels, both manually validated. Validation showed that the NDVI-method achieved median precision 0.51 and recall 0.52, and was more sensitive to phenomena gradually reducing NDVI in actual detection (e.g., multiple shallow tillage). The Color-method demonstrated better overall performance, with median precision 0.84 and recall 0.77. The areas of glyphosate use detected by the two methods were 52,682 and 38,923 ha, respectively. Analysis based on crop type changes and soil types revealed that for Dutch agricultural land in spring, glyphosate is mainly used on sandy soils, to destroy cover crops in cropland, and to remove grass or weeds entirely in grasslands for conversion to cropland. This study demonstrates the potential of remote sensing for quantifying glyphosate use at large spatial scales, making direct detection of glyphosate use possible. However, factors including regional climate and ploughing patterns affect data availability, remaining a limitation.