Neural network clustering for crops thermal mapping

The reliable knowledge of fields, orchards and vineyards, in terms of identification and quantification of crops, plays a key role in precision agriculture. Nowadays, with the recent extension of the unmanned aerial vehicle (UAV) employment in agriculture, a huge amount of high spatial and temporal resolution aerial multispectral imagery is available. In this context, since the temperature of plants can be profitably related to the health status of the crops (e.g. water stress), aerial thermal images can be a valuable tool to extend crop monitoring, which is usually performed locally, to the entire plot and field. The detection of field regions with uniform temperatures is not trivial since aerial imagery is usually characterized by dissimilar patterns at plant scale (e.g. crop rows and shadows). This can deceive standard clustering algorithms into focusing on excessively small-scale thermal differences, losing effectiveness from an agronomical point of view. In this work, a clustering method to detect homogeneous regions in aerial high resolution thermal imagery is presented. The implemented algorithm, based on unsupervised neural networks (NN), does not require a definition of the number of clusters. To demonstrate the effectiveness of the proposed map clustering approach, a vineyard was selected as a case study. A UAV flight with an airborne thermal camera was performed and, contextually, in-field proximal thermal measurements were performed on 72 vineyard regions. The results showed the soundness of the obtained clustered thermal map, confirmed by a positive ANOVA between cluster means and in accordance with the in- field reference measurements. The proposed clustering methodology proved to be robust when faced with sharp temperature gradients, finding crop thermal homogeneity at plot-scale, or at a fraction thereof. Moreover, the clustered map can profitably be embedded in an optimal meteorological sensor network deployment, helping to avoid an inappropriate or redundant placement of nodes.