Search

Amazonian tropical forests play a significant role in global water, carbon and energy cycles. Satellite remote sensing is presented as a feasible means in order to monitor these forests. In particular, the Moderate Resolution Imaging Spectroradiometer (MODIS) is amongst major tools for studying this region. Nevertheless, MODIS operative surface variable retrieval was reported to be impacted by cloud contamination effects. A proper cloud masking is a major consideration in order to ensure accuracy when analysing Amazonian tropical forests current and future status. In the present study, the potential of supervised machine learning algorithms in order to overcome this issue is evaluated. In front of global operative MODIS cloud masking algorithms (MYD35 and the Multi-Angle Implementation of Atmospheric Correction Algorithm (MAIAC)) these algorithms benefit from the fact that they can be optimized to properly represent the local cloud conditions of the region. Models considered were: Gaussian Naïve Bayes (GNB), Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), Random Forests (RF), Support Vector Machine (SVM) and Multilayer Perceptron (MLP). These algorithms are able to provide a continuous measure of cloud masking uncertainty (i.e. a probability estimate of each pixel belonging to clear and cloudy class) and therefore can be used for probabilistic cloud masking. Truth reference dataset (a priori knowledge) requirement was satisfied by considering the collocation of Cloud Profiling Radar (CPR) and Cloud Aerosol Lidar with Orthogonal Polarization (CALIOP) observations with MODIS sensor. Model performance was tested using three independent datasets: 1) collocated CPR/CALIOP and MODIS data, 2) MODIS manually classified images and 3) in-situ ground data. For satellite image and in-situ testing results were additionally compared to current operative MYD35 (version 6.1) and MAIAC cloud masking algorithms. Satellite image and in-situ testing results show that machine learning algorithms are able to improve MODIS operative cloud masking performance over the region. MYD35 and MAIAC tend to underestimate and overestimate the cloud cover over the study region, respectively. Amongst the models considered, probabilistic algorithms (LDA, GNB and in less extent QDA) provided better performance than RF, SVM and MLP machine learning algorithms as they were able to better deal with the viewing conditions limitation that resulted from collocating MODIS and CPR/CALIOP observations. In particular, best performance was obtained for LDA with a difference in Kappa coefficient (model minus MODIS operative algorithm) of 0.293/0.155 (MYD35/MAIAC, respectively) considering satellite image testing validation. Worst performance was obtained for MLP with a difference in Kappa coefficient of 0.175/0.037. For in-situ testing, models overall accuracy (OA) and Kappa coefficient values are higher than MYD35/MAIAC respective values. Models are computationally efficient (swath image calculation time between 0.37 and 9.49 s) and thus being able to be implanted for remote-sensing vegetation retrieval processing chains over the Amazonian tropical forests. LDA stands out as the best candidate because of its maximum accuracy and minimum computational associated.

Amazonian tropical forests play a significant role in global water, carbon and energy cycles. Considering the relevance of this biome and the climate change projections which predict a hotter and drier climate for the region, the monitoring of the vegetation status of these forests becomes of significant importance. In this context, vegetation temperature and evapotranspiration (ET) can be considered as key variables. Vegetation temperature is directly linked with plant physiology. In addition, some studies have shown the existing relationship between this variable and the CO2 absorption capacity and biomass loss of these forests. Evapotranspiration resulting from the combined processes of transpiration and evaporation links the terrestrial water, carbon and surface energy exchanges of these forests. How this variable will response to the changing climate is critical to understand the stability of these forests. Satellite remote sensing is presented as a feasible means in order to provide accurate spatially-distributed estimates of these variables. Nevertheless, the use of satellite passive imagery for analysing this region still has some limitations being of special importance the proper cloud masking of the satellite data which becomes a difficult task due to the continuous cloud cover of the region. Under the light of the aforementioned issues, the present doctoral thesis is aimed at estimating the land surface temperature and evapotranspiration of the Amazonian tropical forests using remote sensing data. In addition, as cloud screening of satellite imagery is a critical step in the processing chain of the previous magnitudes and becomes of special importance for the study region this topic has also been included in this thesis. We have mainly focused on the use of data from the Moderate Resolution Imaging Spectroradiometer (MODIS) which is amongst major tools for studying this region. Regarding the cloud detection topic, the potential of supervised learning algorithms for cloud masking was studied in order to overcome the cloud contamination issue of the current satellite products. Models considered were: Gaussian Naïve Bayes (GNB), Linear Discriminant Analysis (LDA), Quadratic Discriminant Analysis (QDA), Random Forests (RF), Support Vector Machine (SVM) and Multilayer Perceptron (MLP). These algorithms are able to provide a continuous measure of cloud masking uncertainty (i.e. a probability estimate of each pixel belonging to clear and cloudy class) and therefore can be used under the light of a probabilistic approach. Reference dataset (a priori knowledge) requirement was satisfied by considering the collocation of Cloud Profiling Radar (CPR) and Cloud Aerosol Lidar with Orthogonal Polarization (CALIOP) observations with MODIS sensor. Model performance was tested using three independent datasets: 1) collocated CPR/CALIOP and MODIS data, 2) MODIS manually classified images and 3) in-situ ground data. For the case of satellite image and in-situ testing, results were additionally compared to current operative MYD35 (version 6.1) and Multi-Angle Implementation of the Atmospheric Correction (MAIAC) cloud masking algorithms. These results showed that machine learning algorithms were able to improve MODIS operative cloud masking performance over the region. MYD35 and MAIAC tended to underestimate and overestimate the cloud cover, respectively. Amongst the models considered, LDA stood out as the best candidate because of its maximum accuracy (difference in Kappa coefficient of 0.293/0.155 (MYD35 /MAIAC respectively)) and minimum computational associated. Regarding the estimation of land surface temperature (LST), the aim of this study was to generate specific LST products for the Amazonian tropical forests. This goal was accomplished by using a tuned split-window (SW) equation. Validation of the LST products was obtained by direct comparison between LST estimates as derived from the algorithms and two types of different LST observations: in-situ LST (T-based validation) and LST derived from the R-based method. In addition, LST algorithms were validated using independent simulated data. In-situ LST was retrieved from two infrared radiometers (SI-100 and IR-120) and a CNR4 net radiometer, situated at Tambopata test site (12.832 S, 62.282 W) in the Peruvian Amazon. Apart from this, current satellite LST products were also validated and compared to the tuned split-window. Although we have mainly focus on MODIS LST products which derive from three different LST algorithms: split-window, day and night (DN) and Temperature Emissivity Separation (TES), we have also considered the inclusion of the Visible Infrared Imaging Radiometer Suite (VIIRS) sensor. In addition, a first assessment of the Sea and Land Surface Temperature Radiometer (SLSTR) is presented. Validation was performed separately for daytime and nighttime conditions. For MODIS sensor, current LST products showed Root Mean Square Errors (RMSE) in LST estimations between 2 K and 3K for daytime and 1 K and 2 K for nighttime. In the best case (with a restrictive cloud screening) RMSE errors decrease to values below 2K and around 1 K, respectively. The proposed LST showed RMSE values of 1K to approximately 2 K and 0.7-1.5 K (below 1.5 K and below 1 K in the best case) for daytime and nighttime conditions, thus improving current LST MODIS products. This is also in agreement with the R-based validation results, which show a RMSE reduction of 0.7 K to 1.7 K in comparison to MODIS LST products. For the case of VIIRS sensor daytime conditions, VIIRS-TES algorithm provides the best performance with a difference of 0.2 K to around 0.3 K in RMSE regarding the split window algorithm (in the best case it reduces to 0.2 K). All VIIRS LST products considered have RMSE values between 2 K and 3K. At nighttime, however VIIRS-TES is not able to outperform the SW algorithm. A difference of 0.7 K to 0.8 K in RMSE is obtained. Contrary to MODIS and the SW LST products, VIIRS-TES tends to overestimate in-situ LST values. Regarding SLSTR sensor, the L2 product provides a better agreement with in-situ observations than the proposed algorithm (daytime difference in RMSE around 0.6 K and up 0.07 K at nighttime). In the estimation of the ET, we focused on the evaluation of four commonly used remote-sensing based ET models. These were: i) Priestley-Taylor Jet Propulsion Laboratory (PT-JPL), ii) Penman-Monteith MODIS operative parametrization (PM-Mu), iii) Surface Energy Balance System (SEBS), and iv) Satellite Application Facility on Land Surface Analysis (LSASAF). These models were forced using remote-sensing data from MODIS and two ancillary meteorological data sources: i) in-situ data extracted from Large-Scale Biosphere-Atmosphere Experiment (LBA) stations (scenario I), and ii) three reanalysis datasets (scenario II), including Modern-Era Retrospective analysis for Research and Application (MERRA-2), European Centre for Medium-range Weather Forecasts (ECMWF) Re-Analysis-Interim (ERA-Interim), and Global Land Assimilation System (GLDAS-2.1). Performance of algorithms under the two scenarios was validated using in-situ eddy-covariance measurements. For scenario I, PT-JPL provided the best agreement with in-situ ET observations (RMSE = 0.55 mm/day, R = 0.88). Neglecting water canopy evaporation resulted in an underestimation of ET measurements for LSASAF. SEBS performance was similar to that of PT-JPL, nevertheless SEBS estimates were limited by the continuous cloud cover of the region. A physically-based ET gap-filling method was used in order to alleviate this issue. PM-Mu also with a similar performance to PT-JPL tended to overestimate in-situ ET observations. For scenario II, quality assessment of reanalysis input data demonstrated that MERRA-2, ERA-Interim and GLDAS-2.1 contain biases that impact model performance. In particular, biases in radiation inputs were found the main responsible of the observed biases in ET estimates. For the region, MERRA-2 tends to overestimate daily net radiation and incoming solar radiation. ERA-Interim tends to underestimate both variables, and GLDAS-2.1 tends to overestimate daily radiation while underestimating incoming solar radiation. Discrepancies amongst these inputs resulted in large absolute deviations in spatial patterns (deviations greater than 500 mm/year) and temporal patterns.