Data fusion and classification of electromagnetic induction and remote sensing data for management zone delineation in sustainable agriculture 

Abstract

A precise and reliable characterization of intra-field heterogeneity of soil properties and water content is vital in precision agriculture as these significantly impact crop growth and yield. Non-invasive hydrogeophysical methods such as electromagnetic induction (EMI) can be used to delineate intra-field agricultural management zones that represent areas where field characteristics tend to be homogeneous and have similar impact on crops. The combination with additional data sources, for example, remote sensing or yield maps, has the potential to improve the quality of the management zones. However, extracting subsurface information from multiple datasets and for large agricultural fields poses several challenges in data harmonization and analysis. The selection of optimal dataset combinations and the influence of different data products on the creation of management zones have also not been sufficiently investigated. In this study, we present an approach to produce intra-field management zones that combines a) electromagnetic induction (EMI) measurements performed with a CMD Mini-Explorer and a CMD Mini-Explorer Special-Edition (with 3 and 6 coil separation, respectively) and b) normalized difference vegetation index (NDVI) from PlanetScope satellite imagery. The method was tested on a 70-ha field of the PatchCrop experiment in Tempelberg, Brandenburg (Germany). This field is challenging to investigate as it contains 30 small patches of 0.5 ha (72 x 72m) that are managed separately. EMI measurements were collected in three different campaigns in 2022 and 2023 depending on the availability of these small patches. The EMI data were automatically filtered, temperature corrected, and interpolated onto a 1x1 meter resolution grid. Furthermore, EMI measurements were normalized by testing different methodologies (min-max, log, and z-transformation) to reduce the influence of measuring in different periods. Satellite NDVI maps with 3 m resolution for selected years within the period 2019-2023 were obtained from PlanetScope and provided information on crop development over the growing season. For validation, yield maps with 10 m resolution for the period 2011-2019 were available. Both the EMI and the NDVI maps revealed the presence of sub-surface heterogeneities that clearly impact plant productivity, but their patterns did not fully match. To delineate agricultural management zones, ISODATA and K-means clustering algorithms were employed by using a) EMI data, b) NDVI maps, and c) a combination of these datasets. Silhouette and elbow methods were used to identify the optimal number of clusters. The adequacy of the resulting management zones was assessed by comparing them to the available yield maps. The results revealed that a combination of EMI and NDVI datasets could often improve the spatial representation of yield patterns, which confirms the relevance of this method for precision agriculture. Nonetheless, further research is needed to assess the relevance of each dataset and to evaluate the applicability in different regions and contexts.