The agricultural sector is the backbone which supports the livelihoods and the economic development of nations across the globe. In consequence, the need for robust and continuous monitoring of agricultural crops is primordial to face the interlinked challenges of growth rate population, food security and climate change. Synthetic Aperture Radar (SAR) sensors have the powerful imaging capability of operating at almost all weather conditions, independent of day and night illumination. By penetrating through clouds and into the vegetation canopy, the incident radar signal interacts with the structural and dielectric properties of the vegetation and soil, thus providing critical information of the crop state, such as height, biomass, crop yield or leaf structure, which can help devise sustainable agricultural management practices. This is achieved by means of the Polarimetric SAR Interferometry (PolInSAR) technique, which by coherently combining interferometric SAR acquisitions at different polarization states allows for the retrieval of biophysical parameters of the vegetation. In this framework, this thesis focuses on the development of crop monitoring techniques that properly exploit satellite-based PolInSAR data. All the known InSAR and PolInSAR methodologies for this purpose have been analysed. The sensitivity of these data provided by the TanDEM-X bistatic system to both the physical parameters of the scene (height and structure of the plants, moisture and roughness of the soil) and the sensor configuration (polarization modes and observation geometry) is evaluated. The effect of different simplifications made in the physical model of the scene on the crop estimates is assessed. The interferometric sensitivity requirements to monitor a crop scenario are more demanding than others, such as forests. Steep incidences associated with the largest spatial baselines provided by the available data set lead to the most accurate estimates under all the different model assumptions. Shallower incidences, on the other hand, generally yield important errors due to their characteristic shorter spatial baselines. Through the methodologies proposed in this thesis, PolInSAR data have shown potential to refine current methods for the quantitative estimation of crop parameters. Results encourage to continue further research towards the objective of achieving operational crop monitoring applications. / Work supported by the Spanish Ministry of Science and Innovation, the State Agency of Research (AEI) and the European Funds for Regional Development (EFRD) under Project TEC2017-85244-C2-1-P. Noelia Romero-Puig received a grant from the Generalitat Valenciana and the European Social Fund (ESF) [ACIF/2018/204].

For some time now, there has been a growing awareness in society about climate change, pollution, energy and the use of natural resources. This thinking has permeated society, mainly because the extreme natural phenomena that we are experiencing nowadays are no longer outliers in our time series of meteorological records. In this regard, it has been proven that the actual high temperatures are not only unparalleled, but also consistent around the globe which is something that had not happened until now (Neukom et al., 2019). The XX century was a turning point when it comes to the increase of the landuse for crops. In a context where the population doubled, the crop production for food from 1960 to 2010 tripled, helping to reduce the hungry population. When the world’s population is expected to continue to grow up to 9 billion people (Goodfray et al., 2010) by middle XXI century, it is essential to provide ourselves with the necessary tools to maximise crop production by taking advantage of all the resources available under a sustainable point of view. Under this context, all farmers in the European Union (EU) have the possibility to benefit from the Common Agricultural Policy (CAP), which came into force in 1960. The CAP is responsible for the financing of aid to farmers on a cross-compliance basis, based on the declaration of crop types. Traditionally, the authorities have checked the veracity of declarations in person through field inspections, which is clearly a highly inefficient, impractical and very expensive system. However, in 2018 the European Commission drafted an amendment to the CAP (European Commission, 2018), to be implemented in 2020, recommending the establishment of newprocedures for checking declarations, including the use of satellite data from the Copernicus programme or other new technologies. Among the various satellite technologies, Synthetic Aperture Radar (SAR) (Brown,1967; Curlander and McDonough, 1992) has proven the most reliable,as the images are acquired with a constant pass period and they are not subject to cloud problems (as is the case with sensors working in the optical domain) and information can be acquired both day and night. They are based in a SAR microwave sensor installed on a satellite platform with a forward trajectory which offers side-looking imaging geometries. Working in a range between 300 MHz and 30 GHz, the SAR sensor is in charge of emitting electromagnetic pulses and receiving the resulting echoes from the imaged target, which can help retrieve information about its dielectric properties, geometry, orientation, shape, and its behaviour along time. For a given target, the SAR backscattering response σ0 is function of many parameters (Lee and Pottier, 2017; Dobson et al., 1985): wave frequency, polarisation, imaging configuration, roughness, geometrical structure and dielectric properties. This makes the information extraction a major problem, as identical radar responses from two different targets may lead to the same result. To cope with this problem, the main techniques are based on extending the observation space by working with the full diversity of data. Thus, the main axes of SAR data are: • Time • Polarimetry • Interferometry • Frequency. Time series of radar data constitutes a major source of information for the classification of crops and land cover, since it makes it possible to distinguish between classes by their temporal behaviour: some land covers show a uniform response along time (e.g. urban areas), whereas there are others subject to seasonal changes (e.g. crops). It may happen that different crop species give the same radar response at a given time, however, when the time window becomes larger, and consecutive acquisitions are taken over a shorter time span, the more one can detect abrupt changes in the target over a longer time interval. Polarimetry is sensitive to the shape, orientation and the scattering mechanisms of the scatterers (Boerner et al.,1981; Zyl, Zebker, and Elachi, 1987). In that sense, when using different polarisations it is possible to discern better the true nature of the target, as some features may be visible in one polarisation but not in the others. Regarding multi-spectral data, it also constitutes a major source of information which can be exploited for classification purposes. Working with sensors operating at different frequencies, or wavelengths, provides diversity in the size of the elements of the scene to which the radar is sensitive as the radar backscattering will come from elements the size of the wavelength used it. For all of the above, multifrequency data provide complementary information, as each frequency operates and interacts with elements of the same wavelength or longer, and being transparent to all others. In addition, different bands are also associated with different spatial resolutions, so a high-frequency sensor can complement the classification performance of a low-frequency sensor when there are sufficiently small details in the scene that cannot be appreciated with the spatial resolution available at the lower frequency. From all the 4 axes exposed above, Interferometry (Graham, 1974) is without a doubt the least exploited for classification purposes. While polarimetry is sensitive to the scattering mechanisms of the scene by means of the polarisation information, interferometry adds the third dimension by being sensitive to the spatial distribution of the scatterers (Treuhaft et al., 1996). Coherence and phase difference computed between two complex-valued SAR images are the main descriptors of interferometry (Bamler and Hartl, 1998), and together, can be used to derive topographic information, vegetation structure, and deformation (volcanoes, landslides, etc.). For this reason, interferometry is especially suited for classification of covers in which there is vertical distribution of elements, e.g. urban areas and vegetation (forests and crops). Polarimetric interferometric SAR (PolInSAR) (Cloude and Papathanassiou, 1998; Treuhaft and Cloude, 1999), constitutes the next step forward, and is based on the application of interferometry to all polarisation channels. Polarimetry can identify the different scattering mechanisms in the scene by using the polarisation information, whilst interferometry is able to locate the effective scattering phase centres, which are mainly dependent on frequency, the polarisation employed, the physical, geometrical structure and orientation of the scatterer. By using the combination of both we can retrieve the vertical structure of the scene, which shows a great potential for classification purposes, since classes characterised by similar backscattering or polarimetric responses can be separated if their heights are different (e.g. types of buildings, forests, crops, etc.), whereas classes with similar heights, and hence similar interferometric coherence values (e.g. grass, crops, bare soil, etc.) can be resolved using their polarimetric response. In summary, PolInSAR-based classification is attractive since polarimetric ambiguities are resolved by interferometric information and vice-versa. The lack of exploitation of the 4 data axes in the literature, plus the arrival of a new generation of SAR sensors in the near future such as ROSE-L, BIOMASS and NISAR among others, offers a new range of possibilities in terms of new types of features for classification whose results and impact must be analysed. In this context, there are many types of SAR data (i.e. features) that have not been used yet, acquired from different sensors (Sentinel-1, PAZ, TanDEMX, TerraSAR-X and ALOS-2), and whose diversity axes, either used individually or jointly, have not yet been explored for classification applications. Therefore, the exploration of these new types of SAR data, whose contribution to classification is unknown regarding crop-type mapping, is the main objective of this doctoral thesis, and consequently also its main novelty. Based on the current state of the art of the research topic the main objective of this PhD thesis is to explore the added value of new SAR features, and their potential, alone or used together, for crop type and land cover classification. In the end, several experiments will be carried out, in different test sites, in which the proposed new features will be evaluated and compared with the traditional observables used so far, with the aim of evaluating their internal potential in classification applications. / Work supported by the Spanish Ministry of Science and Innovation, the State Agency of Research (AEI) and the European Funds for Regional Development (EFRD) under Projects TEC2017-85244-C2-1-P and PID2020-117303GB-C22. Mario Busquier received a grant from the University of Alicante UAFPU20-08.

京都大学 / 0048 / 新制・課程博士 / 博士(農学) / 甲第21801号 / 農博第2314号 / 新制||農||1065(附属図書館) / 学位論文||H31||N5173(農学部図書室) / 京都大学大学院農学研究科森林科学専攻 / (主査)教授 神﨑 護, 教授 北島 薫, 教授 大澤 晃 / 学位規則第4条第1項該当 / Doctor of Agricultural Science / Kyoto University / DFAM

The efficient and timely monitoring of forest dynamics is of paramount importance and requires accurate, high-resolution and time-tagged predictions at global scale. Despite numerous methodologies have been proposed in the literature, existing approaches often compromise on accuracy, resolution, temporal fidelity or coverage. To tackle these challenges and limitations, the main objective of this doctoral thesis is the investigation of the potential of artificial intelligence (AI) for the regression of bio-physical forest parameters from spaceborne Earth Observation (EO) data. This work explores for the first time the combined use of TanDEM-X single-pass interferometric products and convolutional neural networks for canopy height estimation at country scale. To achieve this, a novel deep learning framework is proposed, leveraging the capability of deep neural networks to effectively capture the complex spatial relationships between forest properties and satellite data, as well as ensuring the adaptability to different environmental conditions. The design and the understanding of the model is driven by explainable AI principles and by considerations on large-scale forest dynamics, with a great emphasis set on the challenges related to the variable acquisition geometry of the TanDEM-X mission, and by relying on the use of LVIS-derived LiDAR measurements as reference data. Moreover, several investigations are conducted on the adaptability of the developed framework for transferring knowledge to related domains, such as digital terrain model regression and above-ground biomass density estimation. Finally, the capability of the proposed approach to be extended to the use of other EO sensors is also evaluated, with a particular emphasis on the ESA Sentinel-1 and Sentinel-2 missions. The developed deep learning framework sets a solid groundwork for the generation of large-scale products of bio-physical forest parameters from spaceborne EO data. The approach achieves cutting-edge performance, significantly advancing the current state of forest assessment and monitoring technologies.