Spatio-temporal analysis of agricultural landscapeiImages: a superpixel-based approach
- Autores: Angel Mario García Pedrero
- Directores de la Tesis: Consuelo Gonzalo Martín (dir. tes.), Mario Lillo Saavedra (codir. tes.)
- Lectura: En la Universidad Politécnica de Madrid ( España ) en 2016
- Idioma: español
- Tribunal Calificador de la Tesis: Ferrán Marqués Acosta (presid.), Ernestina Menasalvas (secret.), Francisco Javier Marcelo Ruiz (voc.), Dionisio Rodríguez Esparragón (voc.), Jan Dirk Wegner (voc.)
- Programa de doctorado: Programa Oficial de Doctorado en Computación Avanzada para Ciencias e Ingenierías
- Materias:
- Enlaces
- Tesis en acceso abierto en: Archivo Digital UPM
- Resumen
In the last hundred years, the world population has tripled and is still growing dramatically butresources have remained the same, causing changes in food supply outlook. According to theFood and Agriculture Organization (FAO), global food production will need to grow by 70% inorder to satisfy the food and feed demand of a population of 9 billion people by 2050. In thecurrent scenario, with limited arable lands and water scarcity, this means a greater pressurethan ever before on productive land; thus today’s main challenge is to put agriculture on a moresustainable and productive long-term path. In this context, the study of the variability withinplots represents an opportunity for improving agricultural management. For most farmers,agricultural plots are variable which implies that not all areas within them require the samemanagement; for example: some areas have specific needs of water, fertilizers, or pesticides inorder to be more profitable. Therefore, adjusting the management of the plots based on thevariability within them opens a route to a variable farm management able to improve cost/yield relationship.
New generation of optical remote sensors placed on aircraft, satellite platforms and drones,offers accessible and useful data of very-high resolution for monitoring and determining spatialvariability of agricultural fields at plot level. However, the landscape complexity makes themanual analysis of the variability within an agricultural plot a highly time-consuming andexpensive task. At the same time, it hinders considerably the monitoring of a high numberof agricultural plots simultaneously. Moreover this very-high spatial resolution represents achallenge for traditional approaches of analysis based on pixels, unable to handle the within-class spectral variability, intrinsic to this type of images. Therefore, it exists a critical need todevelop methodologies for efficiently and automatically extracting and analyzing informationfrom very-high resolution images, that would allow stakeholders to enable variable farmmanagement. In this context, it is of particular interest to automatically detect region dynamicswithin the agricultural plots.
The aim of this work is the development of a methodology for automatic generation of spatialand temporal information on the dynamics of agricultural land, particularly at plot level. In thissense, the study of agricultural scenes has been addressed through an object-based approach,exploring superpixel methods, which are seen as a link between the pixels of the image andobjects of interest. To this end a superpixel method for multi-spectral images has been proposed.This method has been exploited in different applications at three different scales of interest: (i) to analyze fragmented agricultural scenes, (ii) to delineate agricultural plots, and (iii) tocapture the internal variability of agricultural plots.
The analysis of fragmented agricultural scenes has been approached by two methodologies.The first one focuses on providing a framework for multi-scale segmentation and, at the sametime, a way to identify the best scale according to criteria of spectral variability of the regionsat each scale. While this methodology provides spectrally similar regions at all scales, thevariability of agricultural covers hinders to correctly establish the plot boundaries. This issue isfurther discussed in later approaches. The second methodology is aimed at generating thematicmaps by combining only two scales which correspond to segments generated by a segmentationbased on edges and the other based on superpixels. For mapping lands covers, it has combinedthe results of classifying the superpixels through a supervised classifier and the segments basedon edges by a set of rules. The combination of both scales has yielded results with an accuracybetter than those obtained on both scales separately.
Delineation of agricultural plots has been approached by two different methodologies. Thefirst approach uses a supervised classification method to segment the image by agglomeratingsuperpixels. This methodology represents an alternative to traditional methods of segmentation,which is based on learning how agricultural plots in a similar way as a human operator, as itdoes. In this regard, unlike conventional methods that require to find a suitable combination ofsegmentation parameters, the proposed method has the advantage that the classifier is able tofind relationships in a multidimensional space to facilitate an adequate segmentation. The maindrawback of this method is that it requires large amounts of annotated information for trainingthe classifier. In this regard, a second method which uses only the image information for thedelimitation has been proposed. From the basis that segmentations (superpixels) obtainedby different parameters can capture various phenomena at different scales of an image, theproposed methodology allows to exploit its edges to obtain by consensus a segmentation ofagricultural plots. It was found that superpixels can reduce the image noise, while most ofthe edges corresponding to the plots are kept in different segmentations. Both approaches arecomplementary, and depending on the availability of annotated data one or the other can beused, the first method can be used when there is ground truth, and the second one when thereis no such information.
Finally, to capture the internal variability of agricultural parcels, it has proposed a methodbased on superpixels that considers spatial and temporal components of multiple images to findtemporarily homogeneous regions within plots. Further analysis of the behavior of generatedsuperpixels through the different dates provided the information about the spatio-temporalvariability inside the plots. In this regard, three behaviors have been considered to determinethe variability: (i) no variability, (ii) low-persistence variability, and (iii) high-persistencevariability. This considerations allowed to create a map of spatio-temporal dynamics thatprovides an overview of the variability inside plots.
