Identificación basada en objetos de cultivos hortícolas bajo invernadero a partir de imágenes ópticas de satélite

Resum

Agriculture under plastec covered greenhoses represent a step forward in the evolution from traditional to industrual farming. However. this agricultural model has been also criticized for its associated enviromental impacts suchs as plastic residue, soil pollution, viusal impact, biodiversity degradation and local runoff alteration. In this sense, timely and effective greenhouses mappong is the only way to help for yield estimation at different scales, sustainable crop production, estimation and management of residue and the visual and environmental impact in general. Conventional approaches based on in-situ surveys, which are costly and time consuming, are being replaced by supervised classificación of digital information (features) extracteed for satellite images. There are many researches that have been studying the monitoring and mapping of greenhouses using a sort of satellite images and differtent approaches for classificatión, trying to reach the best classification accuracy in the final products. However, few of them have gone a step further and have tried to classify the under greenhose crops, mainly due to the big technical challenge that this spatio-temporal monitoring system entail. In this doctoral thesis we intented to face the callenge of completely remotely classifying under greenhouse crops from time series of medium and very high resolution satellite images by applying an approach based on Object Based Image analysis (OBIA). In this way, we first devise a methodology for the extraction of three-dimensional ground control points from auxiliary free available data of global coverage in order to guarantee the correct georeferencing of the images used. We also propose a workflow to produce high value-added 3D geospantial products such a s Digital Surface Models (DSM) and Digital Terrain Models (DTM) from very high resolution (VHR) satellite images over greenhoses areas. This 3D information layer could be very valuable, as a complement to the traditional 2D spectral data, to improve large-scale mapping of greenhouses. Within the segmentation stage, the command line tool AsseSeg®, coded in Python to carry out a supervised evaluation of the digital image segmentation quality. is development by implementing a modified version of the supervised discrepancy measure called Euclidean Distance 2 (ED2). The next step was the binary classification of greenhouses (i.e., greenhouse vs non-greenhouse) in an OBIA environment throughout the combined use very high resolution satellite data (WorldView-2), Landsat-8 Operational Land Imager (OLI) and Sentinel-2 Multi Spectral Instrument (MSI) and using decision tree classifier. Once the workflow for the classification of greenhouses has benn devided, the callenge of classifying the crops that they house is faced. As in the previous phase (i.e., binary classification of greenhouses), robust decision trees are built from several indices extracted from multitemporal satellite imagery and applied to classify the objects previously segmented (OBIA approach). The results obtained are very promising, reaching overall classification accuracies ranging from 92% to 94% in the case of greenhouses classification, while they were ranging between 74% and 76% in the case of under greenhouse crops classification, turning to be very variable depending on the kind of crops, presenting Fβ value above 95%. The Fβ value for pepper ranged between 66% and 87%, while in the case of tomato the value Fβ ranged between 68% and 75%.