Mineração de dados espaciais aplicada no delineamento de unidades de gestão diferenciada em agricultura de precisão

Abstract

Precision Agriculture (PA) is an agricultural cultivation strategy which uses technologies and principles to manage the spatial and temporal variability related to all aspects that surround a crop, in order to increase yield in a sustainable way, enabling both the reduction of environmental risks and the increase of profits. One of the processes used by this strategy to achieve this goal is the delineation of the crop area in smaller plots with similar characteristics, known as differentiated management units (DMUs). In order to achieve success in this process, dissimilarities between the DMUs must be properly identified from spatial data collected through field or remote sensors. Therefore, the delineation of DMUs may be considered as a spatial data clustering oriented process, in which a clustering solution corresponds to a map of DMUs. The computational approaches found in literature in order to assist in automating this process, usually based in fuzzy clustering algorithms, do not consider, for the most part, the geographic coordinates which compose the collected samples during their methods execution, which can make the DMU maps overly stratified. Therefore, it is possible to observe the lack of a consensus approach in the literature which may allow expert users to obtain DMU maps with minimum internal variability and which, at the same time, are easily interpreted by expert users. Given the above, the process for the delineation of DMUs in PA is discussed in this thesis, whose main contributions are: (i) the SD-Spatial internal validation criteria, which considers issues related to clusters cohesion and separation both in the attribute space and in the coordinate space; (ii) the SWMU Clustering spatial clustering approach, which explores in a weighted way the dissimilarities of the conventional and spatial attributes, using parameters provided by the expert user without the determinism of the solutions being impaired; and (iii) the complementary approach SWMU Polygon, which allows to represent the DMU maps in polygonal shape. Based on the experiments, the SWMU Clustering approach presented average gains of 31.94% in the validation measure considering both the attribute space and the coordinate space, in comparison to approaches using fuzzy clustering; and the complementary approach SWMU Polygon provided average performance gains of 61.14% in the retrieval of DMU maps stored in spatial databases.