Periodic Reporting for period 2 - CALCHAS (Computational Intelligence for Multi-Source Remote Sensing Data Analytics)
Reporting period: 2021-05-01 to 2022-04-30
Climate change is having a profound impact in terms of human health and welfare, contributing to ecological collapse and the destruction of habitats.
Understanding these phenomena necessitates the persistent monitoring of the Earth through the Essential Climate Variables (ECV) which include soil moisture, land surface temperature, and land cover.
The field of Earth Observation (EO) is undergoing an unprecedented revolution in terms of the amount of data collected, facilitating the large-scale monitoring of the environment.
Unfortunately, the increase in volume, variety, and complexity of measurements has led to a situation where data analysis is causing a bottleneck in the observation-to-knowledge pipeline.
To address this challenge, innovative Machine Learning (ML) paradigms like deep neural network models, have established themselves as fundamental tools for EO data analytics, successfully tackling problems such as scene classification and object detection.
Despite the success of these approaches, there is a clear need for establishing new paradigms which will address major challenges in ML-enabled EO data analytics.
These challenges include integrating observations from multiple sources and modalities, addressing the diversity between spaceborne (tens of kilometers) and in-situ sampling scales, as well as analyzing time-series of dynamic observations.
Within the CALCHAS project, we developed an innovative signal processing and machine learning framework that offered the ability to jointly analyze long sequences of EO measurements from satellite and on-ground (in-situ) sensors, achieving high resolution and accurate estimation of critical geophysical parameters.
Our experimental analysis demonstrated that the proposed approach achieved an x2 reduction in retrieval error at x9 finer spatial scales compared to gold-standard NASA products.
Furthermore, our models also demonstrated the ability to forecast ECVs, surpassing the performance of state-of-the-art methods by more than 10% in prediction error reduction.
The framework was also considered for detecting flooded regions by modeling changes between land/permanent water and flooded water and achieved 90% accuracy in automatically detecting flooding at 10-meter spatial resolution.
Understanding these phenomena necessitates the persistent monitoring of the Earth through the Essential Climate Variables (ECV) which include soil moisture, land surface temperature, and land cover.
The field of Earth Observation (EO) is undergoing an unprecedented revolution in terms of the amount of data collected, facilitating the large-scale monitoring of the environment.
Unfortunately, the increase in volume, variety, and complexity of measurements has led to a situation where data analysis is causing a bottleneck in the observation-to-knowledge pipeline.
To address this challenge, innovative Machine Learning (ML) paradigms like deep neural network models, have established themselves as fundamental tools for EO data analytics, successfully tackling problems such as scene classification and object detection.
Despite the success of these approaches, there is a clear need for establishing new paradigms which will address major challenges in ML-enabled EO data analytics.
These challenges include integrating observations from multiple sources and modalities, addressing the diversity between spaceborne (tens of kilometers) and in-situ sampling scales, as well as analyzing time-series of dynamic observations.
Within the CALCHAS project, we developed an innovative signal processing and machine learning framework that offered the ability to jointly analyze long sequences of EO measurements from satellite and on-ground (in-situ) sensors, achieving high resolution and accurate estimation of critical geophysical parameters.
Our experimental analysis demonstrated that the proposed approach achieved an x2 reduction in retrieval error at x9 finer spatial scales compared to gold-standard NASA products.
Furthermore, our models also demonstrated the ability to forecast ECVs, surpassing the performance of state-of-the-art methods by more than 10% in prediction error reduction.
The framework was also considered for detecting flooded regions by modeling changes between land/permanent water and flooded water and achieved 90% accuracy in automatically detecting flooding at 10-meter spatial resolution.
CALCHAS’ primary focus was on the introduction of state-of-the-art deep machine learning methods for the analysis of Earth Observation data, specifically:
Soil Moisture Retrieval from Satellite observations where the objective was the high spatial and temporal resolution estimation of surface soil moisture from remote sensing observation, utilizing available in-situ measurements. The proposed model achieves superior accuracy at finer spatial scales compared to the gold-standard NASA products.
Flood Detection from Satellite observations where we demonstrated accurate flood detection and delineation from diverse sets of satellite observations. The developed model can detect flooded areas at different spatial locations, periods, and specific types of available observations.
Forecasting of Essential Climate Variables using climate reanalysis data, focusing on soil moisture and surface temperature. The proposed approach outperformed state-of-the-art methods when evaluated on predicting monthly averages at 1km spatial resolution.
Modeling high-dimensional Signals, where we investigated how to simultaneously encode observations from multiple instruments across time at different spatial locations. The developed model achieved more than 7% increase in land cover classification accuracy compared to state-of-the-art models.
Analyzing large-scale observations by training deep learning models on large-scale multi-source Earth observations over multiple interconnected computing platforms. The developed model demonstrated increased efficiency in resource utilization such as CPU time and network load.
In terms of dissemination & exploitation activities, quantifiable outcomes of the work include:
*The publication of six articles in prestigious journals and the presentation of this work in major international conferences like IGARSS and ICASSP, as well as more specialized ones like the ESA-ECMWF workshop on Machine Learning for Earth System Observation and Prediction.
*The organization of the Earth Data Workshop, which featured speakers from the regional government, academics from different research institutes, and industry. Furthermore, of 3 Ph.D. students and two MSc were co-supervised within the project.
*The funding of three follow-up projectts, two from ESA and one from NASA, as well as the involvement in the NASA active research mission Cyclone Global Navigation Satellite System.
*Participation in dissemination actions including the Researchers Night and the Science is Wonderful event, researching more than 50 students from elementary schools. Furthermore, relevant videos were uploaded on social media platforms, such as YouTube, which have been seen by more than 100 users.
*The development of three open-access Analysis Ready Datasets that include the first Soil Moisture dataset, a large-scale multi-source flood detection dataset, and a climate variable forecasting dataset.
Soil Moisture Retrieval from Satellite observations where the objective was the high spatial and temporal resolution estimation of surface soil moisture from remote sensing observation, utilizing available in-situ measurements. The proposed model achieves superior accuracy at finer spatial scales compared to the gold-standard NASA products.
Flood Detection from Satellite observations where we demonstrated accurate flood detection and delineation from diverse sets of satellite observations. The developed model can detect flooded areas at different spatial locations, periods, and specific types of available observations.
Forecasting of Essential Climate Variables using climate reanalysis data, focusing on soil moisture and surface temperature. The proposed approach outperformed state-of-the-art methods when evaluated on predicting monthly averages at 1km spatial resolution.
Modeling high-dimensional Signals, where we investigated how to simultaneously encode observations from multiple instruments across time at different spatial locations. The developed model achieved more than 7% increase in land cover classification accuracy compared to state-of-the-art models.
Analyzing large-scale observations by training deep learning models on large-scale multi-source Earth observations over multiple interconnected computing platforms. The developed model demonstrated increased efficiency in resource utilization such as CPU time and network load.
In terms of dissemination & exploitation activities, quantifiable outcomes of the work include:
*The publication of six articles in prestigious journals and the presentation of this work in major international conferences like IGARSS and ICASSP, as well as more specialized ones like the ESA-ECMWF workshop on Machine Learning for Earth System Observation and Prediction.
*The organization of the Earth Data Workshop, which featured speakers from the regional government, academics from different research institutes, and industry. Furthermore, of 3 Ph.D. students and two MSc were co-supervised within the project.
*The funding of three follow-up projectts, two from ESA and one from NASA, as well as the involvement in the NASA active research mission Cyclone Global Navigation Satellite System.
*Participation in dissemination actions including the Researchers Night and the Science is Wonderful event, researching more than 50 students from elementary schools. Furthermore, relevant videos were uploaded on social media platforms, such as YouTube, which have been seen by more than 100 users.
*The development of three open-access Analysis Ready Datasets that include the first Soil Moisture dataset, a large-scale multi-source flood detection dataset, and a climate variable forecasting dataset.
Models driven by physical processes have demonstrated exceptional capabilities in estimating the value of Essential Climate Variables (ECVs) like soil moisture and temperature when considering observations from individual platforms.
While extremely beneficial for global-scale models, the spatial and temporal resolution of existing approaches cannot support fine-scale regional monitoring. Within CALCHAS, we demonstrated how data-driven approaches can significantly increase spatial and temporal resolution while offering predictions consistent with physics-based models.
Our research demonstrated that high-quality predictions can be made by training ML models using historical observations, significantly extending the capabilities of existing forecasting models that are based on numerical weather simulations.
While such models are extremely accurate, they are limited in spatial and temporal resolution due to the massive computational complexity of the models. The proposed scheme can be integrated to provide high spatial and temporal resolution estimations.
When these ECVs exceed critical limits, extreme climate events like flooding and heat waves are significantly more likely to occur. Currently, the Copernicus Emergency Management Services focuses on providing rapid services when an emergency activation is triggered, a process that can take from a few hours to numerous days depending on the type of analysis that is required. Within the project, we demonstrated how advanced data-driven models can be developed for the analysis of satellite observations and the automatic delineation of flooded areas at fine spatial scales using observations from satellites that are part of the EU Copernicus program.
Estimating land cover is critical for gaining a better understanding of different processes. A prolific example of such a data source is the Corine Land Cover maps that have been produced in the context of the EU Copernicus program. These maps are produced every six years and characterize the land at 100 meters resolution. Within CALCHAS, we demonstrated that advanced algorithms can automate the analysis of land cover estimation by analyzing long temporal sequences of observations from multiple spaceborne platforms.
While extremely beneficial for global-scale models, the spatial and temporal resolution of existing approaches cannot support fine-scale regional monitoring. Within CALCHAS, we demonstrated how data-driven approaches can significantly increase spatial and temporal resolution while offering predictions consistent with physics-based models.
Our research demonstrated that high-quality predictions can be made by training ML models using historical observations, significantly extending the capabilities of existing forecasting models that are based on numerical weather simulations.
While such models are extremely accurate, they are limited in spatial and temporal resolution due to the massive computational complexity of the models. The proposed scheme can be integrated to provide high spatial and temporal resolution estimations.
When these ECVs exceed critical limits, extreme climate events like flooding and heat waves are significantly more likely to occur. Currently, the Copernicus Emergency Management Services focuses on providing rapid services when an emergency activation is triggered, a process that can take from a few hours to numerous days depending on the type of analysis that is required. Within the project, we demonstrated how advanced data-driven models can be developed for the analysis of satellite observations and the automatic delineation of flooded areas at fine spatial scales using observations from satellites that are part of the EU Copernicus program.
Estimating land cover is critical for gaining a better understanding of different processes. A prolific example of such a data source is the Corine Land Cover maps that have been produced in the context of the EU Copernicus program. These maps are produced every six years and characterize the land at 100 meters resolution. Within CALCHAS, we demonstrated that advanced algorithms can automate the analysis of land cover estimation by analyzing long temporal sequences of observations from multiple spaceborne platforms.