Soil properties zoning of agricultural fields based on a climate-driven spatial clustering of remote sensing time series data

The identification of zones within an agricultural field that respond differently to environmental factors and agronomic management is a key requirement for the adoption of more precise and sustainable agricultural practices. Several approaches based on spatial clustering methods applied to different data sources, e.g. yield maps, proximal sensors and soil surveys, have been proposed in the last decades. The current availability of a huge amount of free remote sensing data allows to apply these approaches to agricultural areas where ground or proximal data are not available. However, in order to provide useful agronomic management information, it is essential that the zoning obtained by clustering is linked to the underlying spatial variability of soil properties. In this work we explore the hypothesis that the response of crop vigor to temporal climate variability, assessed by remote sensing data time series, selected to correspond to specific growth phases and seasonal climate patterns, provides indications on the variability of soil properties within agricultural fields, for both herbaceous and tree crops. NDVI time-series for 38 years (1984–2021) were obtained for fourteen non-irrigated herbaceous and tree crop fields in Central Italy, from multispectral satellites data (Landsat 5/7/8, Sentinel 2). The Standardized Precipitation-Evapotranspiration Index (SPEI) was used to classify time series into three climatic classes (dry/ normal/wet) for five different periods of the growth season, covering the main phenological phases. K-means clustering was used to identify patterns of crop growth from climatically classified image sets, as well as for all the bulked images for comparison (bulk clustering). Clustering results were compared with soil maps obtained from spatialized ground data, for soil texture (clay, silt and sand), soil organic matter and available soil water (ASW). The agreement between the different clustering results and soil maps was assessed by the Adjusted Rand Index. Agreement with soil maps varied depending on the field, the phenological phase considered and the soil property considered. Climate driven clustering from long, late growth season periods best matched soil properties, both for herbaceous and tree crops, despite being based on a limited number of images. The clustering from images spanning a longer growth period for dry years systematically outpaced the bulk clustering for silt, sand and ASW, while the clustering for normal climatic conditions was the best for organic matter. The performance of the matching between clustering and soil maps increased with soil variability significantly more (P < 0.05) than in the bulk clustering (mean slopes respectively 0.468 ± 0.167; 0.113 ± 0.270). The integration of the SPEI climatic index into the clustering procedure systematically improved the identification of zones with homogeneous soil properties, highlighting that a greater attention should be posed to the climate-crop-field interactions when using remotely sensed images