The pool of dissolved organic carbon (DOC) is a pivotal influencer in the ecology and biogeochemistry of inland water ecosystems, constituting a significant factor in the carbon budgets of terrestrial ecosystems. Employing a machine learning approach and leveraging a large curated dataset (AquaSat), this study delves into the importance of multiple spectral remote sensing data and spatiotemporal information in contributing to the variability of DOC concentrations in inland water bodies. The research underscores the critical role of spatial and temporal information in enhancing the accuracy of DOC predictions, revealing a remarkable 50% decrease in RMSE and MAE and a 110% improvement in R2in machine learning model performance when spatiotemporal covariates are incorporated. This marks a significant improvement over the reported accuracy on the same DOC dataset [1]. The dataset’s extensive coverage of inland water bodies provides valuable insights that are crucial for effective water quality monitoring, environmental management, and ecosystem health assessment.

Excessive levels of carbon dioxide (CO2) in the atmosphere contributes to global temperature rise, and efforts are being made to limit this increase to ensure the safety of Earth’s inhabitants. Satellites like GOSAT-2 and OCO-2 provide global-scale monitoring of atmospheric CO2levels. However, cloud and aerosol occlusion result in missing data, and the spatial and temporal resolutions of these measurements are coarse. Addressing these limitations is crucial for leveraging satellite-based global CO2monitoring to identify CO2sources and sinks and understand their spatio-temporal evolution. In this study, we employ machine learning techniques to estimate column-averaged CO2(XCO2) from level 2 (L2) XCO2estimates obtained from the OCO-2 satellite, and daily XCO2data generated using Fixed Rank Krigging (FRK) at a spatial resolution of 10x10is used as the target variable. Meteorological variables, known to strongly influence XCO2distribution, are considered as covariates in the machine learning framework. To validate our estimates, we compare them with measurements from the Total Carbon Column Observing Network (TCCON) sensors. Additionally, we compare our estimates with those obtained from FRK and GEOS-L3 data. The validation against TCCON measurements and the comparison with existing data sources contribute to the evaluation and reliability of our approach.

Indiscriminate use of chemicals for farming leads to various environmental pollution- air, land, and water. Judicious use of chemicals in farming has a huge environmental benefits as well as reduction in farming cost. In this paper, we proposed a novel spatio-temporal adaptation techniques to localize the pest/disease risk using multi modal data - geo-spatial location, weather, satellite observations, and plant imageries. To illustrate the efficacy of the method, we demonstrated a case study with potato farmers which fulfilled the dual objectives of positively bringing food security and safety to our society while enabling sustainable and profitable operations.

Greenhouse gas (GHG) emissions play a significant role in climate change and its adverse impacts on the environment. Accurate estimation of these emissions is crucial for developing effective mitigation strategies. In recent years, advancements in machine learning techniques have opened up new opportunities to improve GHG estimation methodologies by leveraging diverse data sources and modeling approaches. This research presents an innovative ensemble approach that combines satellite data, process-based models, and static calculators to estimate GHG emissions more accurately. The proposed framework harnesses the strengths of each component, resulting in a comprehensive and robust estimation system. The ensemble approach employs machine learning algorithms to integrate and harmonize the outputs from the satellite data, process-based models, and static calculators. The proposed methodology significantly reduces the estimation error to 11.25% compared to the 36.3%, 53.5%, and 17.2% estimation error when satellite, static calculator and process based models were used for estimating the GHG emissions.

Green house gas (GHG) satellites such as Orbiting Carbon Observatory 2 (OCO2), Sentinel 5P (TROPOMI), GHGSat offer global coverage of measuring column averaged carbon-di-oxide (XCO2) and methane (XCH4). Though GHG satellites are a scalable method of measuring GHG emissions, they are limited by coarse spatial and/or temporal resolutions and missing data. Based on the statistical interpolation technique, fixed rank kriging (FRK), in this work, we have developed a novel nested kriging approach, N-FRK to address the above challenges of GHG satellite data, in particular of OCO2 satellite. Compared to the spatiotemporal resolution of 1.2×2.2km2 and 16 days of OCO2 satellite, we have increased the temporal resolution to 1 day and spatial resolution to 1.11×1.11km2using N-FRK. The daily spatial maps at 111, 11.1, 1.11 km resolutions have been generated using FRK and N-FRK techniques and validated across 13 or subset of the 13 sensor sites of total carbon column observing network (TCCON) for the year 2019. As part of validation, we present the R2, root mean square error (RMSE), and bias metrics and we see good agreement between the estimated data and sensor data.

Automated crop classification and mapping is currently a topic of significant research interest worldwide due to the following two factors. First, it is one of the key tasks on which the success of digital agriculture hinges, and second, there is wider availability of remote-sensed imagery, both optical and radar-based, that can help with remote monitoring of crops. Several different models have been developed for the purpose, but a hitherto unexplored geography or time period generally requires fresh ground data specific to the space and time, and, in many cases, fresh feature engineering as well, due to lack of intra-class compactness and inter-class separability. A near-universal model that can be applied with minimal fine-tuning using free and open access satellite imagery and requiring no new ground data, which can reduce data costs, is elusive. This paper underscores the challenges involved via a case study with two different classification approaches at various regions. Methods and techniques that can ameliorate the problems are discussed.

There have been several efforts to utilize satellite-based synthetic aperture radar (SAR) measurements to determine surface soil moisture conditions of agricultural regions. The results have been mixed since the relation between the SAR signal and surface soil moisture is confounded by variations in topographic features, surface roughness and vegetation density etc. We designed an experiment to investigate SAR based soil moisture retrieval using different machine learning techniques. In addition, a high resolution land surface model customized and deployed for generating soil moisture at 250m resolution using various static and dynamics input data.