The proliferation of machine learning models, architectures, and datasets for Earth observation (EO) continues to rise dramatically. This pattern is expected to continue growing bringing with it an increase in the generation of remote sensing derived data products powered by geospatial artificial intelligence (GeoAI) techniques. Rigorous quality assessments and accuracy analysis needs to be undertaken for the science community to adopt many of these data products for scientific discovery of changes of the Earth’s land surface. While there is existing literature supporting and documenting best practices for the validation of GeoAI data products, the software to support large-scale collaborative validation efforts is limited. In this study we present the design and software implementation of a flexible commodity-based framework for large-scale global to regional validation of GeoAI data products. This framework’s main purpose is to enable, speed up, and optimize the acquisition of validation data for large-scale science projects with support across multiple sensors and spatial resolutions with little to no code. In addition, we present several use cases where this framework has enabled and streamlined the validation of global to regional data products at different spatial resolutions and within different computational platforms.

Quantifying aboveground biomass (AGB) in tropical forests is a challenging but necessary task to support actions to preserve and restore these ecosystems. Recent methods have leveraged the integration of NASA’s spaceborne lidar GEDI with imaging sensors to improve AGB predictions. However, the use of high-resolution (<1 m) stereo images from spaceborne sensors still needs to be explored within these frameworks. The objective of this study was to predict AGB in a tropical forest patch by combining stereo images and GEDI data. A digital surface model (Stereo-DSM) was generated by stereophotogrammetric processing of high-resolution stereo pairs collected from a spaceborne sensor. A canopy height model (Stereo-CHM) was then derived by subtracting an airborne lidar-derived digital terrain model from the Stereo-DSM. Descriptive statistics were calculated from the Stereo-CHM to be used as predictors in the model to predict AGB. We also calculated vegetation indices from the Harmonized Landsat-Sentinel and Sentinel-1 images to compare their relative importance in the model and assess their relationship to the Stereo-CHM metrics. Finally, we trained a Random Forest model using stereogrammetry - derived, multispectral and SAR metrics as features and GEDI’s footprint-level AGB product as a reference. The model to predict AGB yielded performance metrics of r = 0.63, RMSE = 25.38 Mg/ha, and MD = 2.1 Mg/ha. The metrics from the Stereo-CHM were ranked as the most important to the model. This is an indication that these metrics can add important information related to canopy structure to inform GEDI-based models to predict AGB. It is still necessary to evaluate these results considering a variation in canopy cover, topography, and understory vegetation. The findings are important to support advances on the integration of large footprint spaceborne lidar and images with sub-meter spatial resolution to characterize vegetation in tropical forests.

Human-induced land cover land use (LCLU) changes such as agricultural extensification and forest degradation and loss have extensive negative impacts including biodiversity loss, land degradation, and a disruption to ecological services. In Senegal, where people are heavily reliant on dryland agricultural production, climate change and land degradation pose particularly significant threats especially as rapid population growth continues to fuel frequent LCLU change. Considering these challenges, approaches that facilitate increased insight into the spatial and temporal dynamics of land use are needed to implement sustainable land management practices and mitigation strategies. However, difficulties associated with Senegal’s highly variable phenology, sparse woody cover and small, irregular fields necessitate the use of Very High Resolution (VHR; < 3 m spatial resolution) data and modern techniques for modeling land use at sufficient scales.We take advantage of VHR data’s spatial resolution and Sentinel-1’s high temporal resolution by implementing an object-based data fusion strategy to model land use. By generating high resolution vector objects from single-date WorldView imagery and using the corresponding Synthetic Aperture Radar (SAR) time series to train a One-Dimensional Convolutional Neural Network (1D CNN), we can effectively leverage deep learning techniques to extract land use signals from multi-resolution and multi-temporal data in a near-autonomous manner.

The Amhara region, in northwest Ethiopia, has a complex topography and highly fragmented croplands (averaging half a hectare). Mapping such fragmented LCLU areas requires very high-resolution satellite imagery and robust classification methodologies. To this end, we have used multi-temporal very-high-resolution (VHR) WorldView imagery (2 meters) in combination with Convolutional Neural Networks (CNNs), to map land cover classes across the entire Amhara region. This paper presents results from domain adaptation experiments using training data from Senegal to accurately map land cover classes at 2 m resolution in the Amhara region, Ethiopia. The Attention UNet CNNs provided promising results for predicting land cover in Ethiopia imagery using domain adaptation techniques and without the addition of local training labels, with an overall accuracy of 74%. We conclude that promising future research directions exist for transfer learning implementation to finetune our land cover classes with additional model refinement to the Amhara region.

Land cover mapping has been a valuable tool in capturing changes in many developing regions in Africa. Senegal has been a hotspot of change where agricultural activity has rapidly increased. Agriculture in this region is often a complex mosaic of small fields which makes them difficult to classify using conventional land cover mapping methods and coarse-resolution satellite imagery. WorldView (WV) satellites provide very high-resolution imagery that is ideal for semantic segmentation using convolutional neural networks (CNN). In this study, we introduced training strategies that scale up the training data for the U-Net model using 2 m WV-2 and 3 imagery to overcome the challenges of regional mapping with a patchwork of hundreds of images. The proposed strategies increased the number of training data for the U-Net model in three main scenarios, (i) conventional training, (ii) model transfer, and (iii) transfer learning, and we evaluated model generalizability on test sets for two different regions in Senegal. The results showed that models rapidly reached a high level of performance with a limited increase in additional training in conventional and transfer learning strategies. In these two strategies, the U-Net consistently produced >87% average accuracy for trained images and >70% average accuracy for all test images at the final scale level. The research opens opportunities to produce regional land cover maps in West Africa without generating a prohibitively large amount of training data.

Validating land cover classification results from a machine learning model is a vital step in ensuring that further decisions are based on sound and robust results that can be trusted. Calculating pixel-wise validating statistics from a stack of land cover classification results, while computationally trivial for low-resolution imagery with a small spatial footprint, poses a significant challenge for very-high-resolution (VHR) imagery spanning a larger spatial footprint. Here we describe an open-source unified Python framework and workflow for the compositing of VHR imagery based on climatic and spatial information leveraging hardware acceleration. We additionally describe the implementation of per-pixel reduction algorithms which are used to reduce the stacked composite into a robust and accurate composite that is validated.

Circumboreal forests are currently experiencing rapid climate warming which is altering their structure, productivity, and status as a carbon sink. Very high-resolution (VHR; < 2 m) stereo-derived digital surface models are available to monitor these forests, but a similar resolution digital terrain model (DTM) is required to extract information about tree height, which is often used to estimate carbon content. To the best of our knowledge, no openly available VHR DTM currently exists. To address this need, we developed approaches to extract DTMs by filtering VHR stereo point clouds (PCs) in sparse canopies of Alaska. Our evaluation consisted of two stereo processing methods with three PC search radii at six different tree canopy cover (TCC) intervals. We found that VHR DTMs were robust for estimating bare ground at TCC intervals less than 40% with vertical errors <1.6 m using airborne small footprint light detection and ranging (LiDAR) as reference.

Smallholder farms in South East Asia are characterized by small irregular field patterns, dense cloud cover and haze which limits our ability to observe changes in agriculture land-use. Very-high resolution (VHR,30 m) in regions with dense persistent cloud cover and haze from biomass burning.

Commercial very high-resolution (VHR) Earth observing (EO) satellites have grown into constellations with global repeat coverage that can support existing NASA EO missions with stereo and multispectral capabilities. Sub-meter data from these instruments exceeds petabytes per year and the cost for data, storage systems and compute power have all dropped exponentially. Concurrently, through agreements with the National Geospatial-Intelligence Agency, NASA-Goddard Space Flight Center is acquiring VHR EO imagery from DigitalGlobe's WorldView-1, 2, 3 Quickbird-2, GeoEye-1 and IKONOS-2 satellites. To enhance the utility of these data we are developing an Application Program Interface (API) to produce on-demand user defined science ready products to support NASA's EO missions. These enhancements include two primary foci: 1) surface reflectance 1/2° ortho mosaics - multi-temporal 2 m multispectral imagery that can be used to investigate biodiversity, horizontal forest structure, surface water fraction, and land-cover land-use at the human scale; and 2) VHR digital elevation models (DEMs) - derived with the NASA Ames Stereo Pipeline. These enhanced products benefit Earth surface studies on the cryosphere (glacier mass balance, flow rates and snow depth), hydrology (lake/waterbody levels, landslides, subsidence) and the biosphere (vertical forest structure, tree canopy height and cover) among others. Here we present current API capabilities and recent examples of derived products used in NASA Earth Science projects.

NASA Wrangler is a loosely-coupled, event driven, highly parallel data aggregation service designed to take advantage of the elastic resource capabilities of cloud computing. Wrangler automatically collects Earth observational data, climate model outputs, derived remote sensing data products, and historic biophysical data for pre-, active-, and post-wildfire decision making. It is a core service of the RECOVER decision support system, which is providing rapid-response GIS analytic capabilities to state and local government agencies. Wrangler reduces to minutes the time needed to assemble and deliver crucial wildfire-related data.