Mapeado de invernaderos de plástico mediante análisis de imagen basado en objetos y series temporales de imágenes de satélite

Resumo

The world population will be about 9.7 billion people by 2050. This population growth will be accompanied by an increase in the demand for products from agriculture. Therefore, one of the most important challenges of contemporary agriculture is to increase its productivity gains while controlling, at the same time, the agriculture’s environmental footprint. In this sense, greenhouse protected agriculture has proven to be one of the most efficient models in recent decades, managing to increase harvests of horticultural crops. However, the expansive use of plastic in agriculture as greenhouses or mulching crops causes negative effects on the environment. A recent report by the Food and Agriculture Organization (FAO) calls for the need to coordinate good management practices with the aim of curbing the "current disastrous use of plastics across the agricultural sector", which is causing the accumulation of dangerous micro-plastics in the environment, mainly in soils and seas. Nowadays, the numerous and frequent medium, high and very high resolution (VHR; Very High Resolution) optical satellite images available represent an important source of data for analyzing the agricultural sector. In addition, many of these sensors generate free and open access periodic data anywhere in the world (e.g., Sentinel-2 or Landsat 8). For this reason, plastic greenhouse mapping techniques using satellite images are being applied more and more frequently within the field of remote sensing. As a first stage of this thesis, a detailed analysis of the worldwide scientific production carried out in the remote sensing field about agricultural greenhouses and plastic mulched crops throughout the 21st century was carried out. In fact, 107 publications from Scopus were located up to the year 2019, analyzing their characteristics and trends, thus concluding that it is an emerging topic, mainly in countries such as China, Italy, Spain, the United States and Turkey. It is important to point out that during the years 2020 and 2021 another 31 works have been published, demonstrating that the topic dealt with in this thesis continues to attract the interest of numerous researchers around the world. Many of the publications mentioned above use time series of optical satellite images. Considering that the combination of images from Sentinel-2 (including its twin satellites 2A and 2B) and Landsat 8 provides a global median average revisit interval of 2.9 days, it is not surprising that many researchers have used both satellites to generate their time series. In a second stage of the thesis, the spectral consistency of the surface reflectance values of the Sentinel-2 and Landsat 8 images was evaluated over areas of plastic greenhouses located in Spain, Morocco, Italy and Turkey. In general, high correlations were found in most of the bands and in the NDVI vegetation index, although also discrepancies were found. These were mainly due two reasons. (i) The existence of solar reflections on the roofs of the greenhouses related to the orbit of the sensor and the position of the sun at the time of image capture. (ii) The content of aerosols in the atmosphere, causing important differences in the surface reflectance values of Sentinel-2 and Landsat 8, mainly in the visible bands with shorter wavelength. The greenhouse detection approach proposed in this thesis is based on object-based image analysis (OBIA). In this way, we need to start from VHR satellite images to perform the segmentation process, in order to obtain the objects that represent plastic greenhouses. In the third part of the thesis it was decided to test the only Spanish VHR satellite called Deimos-2 (1 m resolution in panchromatic), of which there were no previous publications regarding its capabilities to generate georeferenced products. That is why the third part of the thesis was devoted to studying the level of precision achievable, both under operational conditions and on different land covers (bare land, urban areas and greenhouse zones), of products such as orthoimages and digital surface models generated from Deimos-2 stereo pairs. The obtained results validated the use of Deimos-2 images in our OBIA greenhouse detection process. In the fourth phase of the thesis, a test was carried out with a WorldView-3 image (VHR satellite with 16 bands) taken over the greenhouses of Almería, testing several plastic detection indices that have already been used in remote sensing for classification of plastic greenhouses or marine plastic debris. The NDPI (Normalized Difference Plastic Index) based on WorldView-3 SWIR bands was the one that yielded, by far, the best results in classifying plastic in general, and greenhouses in particular. The last phase of the thesis focused on mapping plastic greenhouses based on OBIA techniques, using Deimos-2 for segmentation and Sentinel-2 time series for classification. Several greenhouse detection indices calculated on Sentinel-2 images and time series were tested on four different greenhouse areas (Almería (Spain), Agadir (Morocco), Bari (Italy) and Antalya (Turkey). This fact gave great strength to the results obtained. The PGHI (Plastic GreenHouse Index) based on the blue band and SWIR2 of Sentinel-2 was the one that obtained the best results with an overall plastic covered greenhouse classification accuracy of around 93%.