Remote sensing to support surveillance, response and eradication: New Zealand maize crops spatial distributions

Introduction: When preparing for and responding to incursions by organisms such as insects, it would be useful to know the distributions of their host plants to inform surveillance and management. We are investigating the use of satellite data for identifying the spatial distributions of several plant species that are hosts for some important biosecurity hazards. The main challenge is to discriminate plants with similar seasonal and trait characteristics at the landscape scale. 

Methods: This study developed a new method for differentiating maize−a plant threatened by numerous biosecurity hazards−from other New Zealand crops based on the Sentinel−2 satellite time series (TSs) method. Combining phenological information from multiple temporal, spatial and spectral bands and the vegetation indexes contained in the open and free images acquired by Sentinel−2 enables seasonal trend analysis, which is extremely useful for crop type mapping. However, a lack of large-scale training datasets hampers development and testing of advanced methods for agricultural phenology-based classifications. To improve the training data, we used the TimeSen2Crop dataset captured in Europe. To capture crop phenology, Sentinel−2 imagery dating from 1st of September to 31st of April for five growing seasons was used. Surface reflectance data was filtered to be cloud−and shadow−free. Different adjustments to the temporal component of the training data were tested to match Southern Hemisphere seasons. Phenological curves were extracted from time series by fitting harmonic regressions, which decomposes signals over time into their constituent frequencies (i.e. cosines and sines). Subsequently, various Machine Learning (ML) models were compared, and Random Forest (RF) was selected for optimal accuracy over a number of seasons and regions. Finally, and for the latest growing season, an opportunity to further improve classification accuracy was explored by incorporating New Zealand data from previous seasons. After semi-automatic maize paddocks vectorization from aerial imagery, matching Sentinel−2 images were transformed into a training data and used for augmentation of the main training dataset. 

Results and Discussion: The Green Chlorophyll Vegetation Index computed as harmonic regression accurately depicts phenological curves. The overall accuracy of classifying maize crop areas reached about 92% based on c. 1400 field samples. The results demonstrate that: (1) landuse and crop type classification based on Sentinel−2 time series can be highly effective for identifying crop phenological information to help optimise surveillance and incursion responses over large areas; (2) surveillance resources can be allocated to the most appropriate locations after suitable sampling sites have been identified