Remote Sensing Image Processing and Classification Techniques

Experts in the field of image analysis and classification will present applications of single and fused data sets for mapping and monitoring vegetation, accuracy assessment considerations, and how these data are used in decision making.

2:00- 2:15 PM – Landsat-Derived Global Rainfed and Irrigated-Cropland Product @ 30 M (LGRIP30) For Conterminous United States (CONUS) For The Year 2020
The global food and water security scenario in the twenty-first century will be an extraordinarily complex one. The population of the world will continue to balloon and reach 9.7 billion by 2050 and nearly 11-12 billion by 2100. In addition, global daily average calorie consumption is expected to rise from 2789 kcal/person/day in the year 2000 to 3130 kcal/person/day by the year 2050. These consumption figures may rise even higher if traditionally low meat consuming nations start increasing meat consumption enabled by economic growth. Food habits of people are diversifying considerably (e.g., rice, wheat, or maize only to a mix of rice, wheat, maize, pulses, fruits, and vegetables). The United Nation’s Food and Agriculture Organization has noted the potential grave danger to global food supply because of war and pandemic related supply-chain issues. Additionally, global food and water security challenges are tightly intertwined. In a changing climate, the quantity and quality of surface and ground water are decreasing. Demands for agricultural and alternative uses (e.g., industrial, ecological) of water are simultaneously increasing. Adding to this complexity, urban migration and climate change related precipitation extremes necessitate new infrastructure to maintain water availability for new cropland. Thus, 30m (field scale where 1 pixel = 0.09 hectares) cropland products, including their intensity and irrigation status, are needed to monitor change and aid decision-making for issues of global food and water security. Importance of mapping irrigated and rainfed croplands cannot be overemphasized. Both irrigated and rainfed areas are water guzzlers since about 80-90% of all human water use goes towards producing food. Irrigated areas consume blue water. That is the water delivered to farms through irrigation systems either from surface water (e.g., lakes, reservoirs, tanks, river diversions) or ground water resources (e.g., tube wells, open wells). Rainfed areas consume green water. That is water coming from direct precipitation (e.g., rainfall, snowfall, soil moisture).   Thereby, the overarching goal of this research was to produce the highest-resolution Landsat-derived global rainfed and irrigated cropland product @ 30m (LGRIP30) using machine learning, and cloud computing. The study will focus on highest known resolution, accurate LGRIP30 product for the Conterminous United States for the year 2020 (LGRIP30 CONUS 2020). The LGRIP30 cropland product was generated using Landsat-8 30m time-series data, multiple supervised and unsupervised Machine learning algorithms (MLAs) such as random forest, support vector machines, decision trees, ISOCLASS clustering, and spectral matching techniques. The study will make use of a fusion of Landsat-8, 9, and Sentinel-2A&2B (S2) surface reflectance (SR) products already available in Google Earth Engine (GEE), and NASA’s Harmonized Landsat Sentinel-2 (HLS) Landsat product (HLSL30) for 2013-present and HLS Sentinel-2 product (HLSS30) for 2015-present, that together have 2-3 day global coverage at nominal 30m resolution. Data for every 16 days from January 1, 2014, through December 31, 2016, was used in processing. In total 10 bands (blue, green, red, NIR, SWIR1, SWIR2, TIR, EVI, NDWI, NDVI) of data were processed. All Landsat-8 images were cloud masked using CFMask in GEE. Other 30m resolution bands such as the Shuttle Radar Topography Mission (SRTM) elevation and slope, soils, and local information were added to data cubes for each AEZ as needed to best help separate classes. These Landsat analysis ready data cubes (ARD-cube) were composed on a cloud platform like GEE to help seamlessly code and compute using MLAs.
Prasad Thenkabail, USGS

2:15 – 2:30 PM – Quantifying Urban Vegetation and Non-Functional Turfgrass Using High-Resolution Aerial Imagery And Lidar Data In Las Vegas, Nevada
In August 2021, following two decades of drought, the US federal government declared Lake Mead in a water shortage for the first time since its construction in the 1936. To mitigate unnecessary water loss, the city of Las Vegas, Nevada has enacted an ambitious piece of conservation legislation prohibiting the use of Colorado River water delivered by the Southern Nevada Water Authority (SNWA) member agencies to irrigate ‘useless’ or non-functional turfgrass not zoned exclusively for single-family residences by 2027. Therefore, all non-functional turfgrass must be removed by the end of 2026. To assist its removal, an accurate estimate of urban vegetation and non-functional turf area and its spatial location is needed. I used airborne lidar and high-resolution 4-band orthophotos to map urban turfgrass, tree and shrub cover in Las Vegas Valley (LVV) then performed spatial analysis to classify non-functional turfgrass. I found that turfgrass, tree, and shrub area totaled 8,331 acres, 21,800 acres, 4,848 acres, respectively. I found that there was 353 acres of non-functional turf in the study area with commercial business parcels, non-profit, government, and religious facility parcels containing the greatest area of non-functional turfgrass and should be targeted for removal.
Megan Singleton, Southern Nevada Water Authority

2:30 – 2:45 PM – High Resolution Salt Marsh Habitat Mapping and Change through Machine Learning
Historically, New England salt marshes have been dominated by high marsh meadows characterized by species such as Spartina patens and Distichlis spicata. In recent decades, annual field surveys have shown an increase in the distribution of Spartina alterniflora in Rhode Island marshes, indicating a shift in the ecosystem due to the impacts of sea level rise. This compositional shift highlighted the need to update distribution maps to quantify the changes occurring in these vulnerable ecosystems. The NOAA Office for Coastal Management, in partnership with the Narragansett Bay National Estuarine Research Reserve, contracted NV5 Geospatial (NV5G) to map critical salt marsh habitat types as well as invasive species such as Phragmites australis. Utilizing high-resolution multispectral aerial imagery and detailed field observations, NV5G built machine learning models to efficiently update the state’s salt marsh habitat maps. These updated maps are now being used by local experts to prioritize areas for restoration and to monitor the effectiveness of invasive species eradication efforts.
Chris Robinson, NV5 Geospatial

2:45 – 3:00 PM – Key Issues When Conducting a Thematic Map Accuracy Assessment
While assessing the accuracy of a thematic map generated from remotely sensed imagery is often a required step in the mapping process, there are still many issues with conducting the assessment correctly.  Some of these issues are related to methodologies imposed by the software used for the mapping project.  Others include statistical considerations such as accounting for spatial autocorrelation, appropriate sample size or sample unit, and sampling strategy.  Failure to consider these issues can result in an invalid assessment.  This presentation demonstrates these important issues and discusses how the informed user can make sure that the accuracy assessment they are conducted is valid.
Russell Congalton, University of New Hampshire

3:00 – 3:15 PM – Data-Driven Flood Risk Index by Country for Downstream Decision-Making
Flooding is one of the most frequent and costliest hydro-meteorological hazards that impacts every country worldwide and contributes to significant societal and financial losses. While there is no shortage of Earth Observation (EO) datasets and hydrodynamic models to map, forecast and monitor flood events, decision makers and first responders face significant challenges in using these products. To assist with resource planning, emergency management and downstream analytics, we have developed a data-driven machine learning model – Flood Risk Index for Resilience (FRI-R). The index is built on Model of Models (MoM), an operational open source ensemble model, that integrates outputs from outputs from hydrologic models and EO data (optical imagery) and forecasts flood severity globally at sub-watershed level every 24-hours. Based on historical MoM outputs, the index accounts for spatial and temporal distribution of flood events and identifies high to low risk watersheds globally based on the duration and frequency of past flood events. Considering the global focus to improve climate resilience of communities, MoM has been successful in disseminating early flood warning based on forecasted risk. FRI-R, however, expands the usability of MoM as it identifies high flood risk watersheds which could be used to geotarget priority populations and critical infrastructures to mitigate flood impacts and improve resilience.
Bandana Kar, U.S. Dept. of Energy

3:15 – 3:30 PM – Radiometric Harmonization using Cubist-based Machine Learning for Creating Multi-sensor Image Mosaics
In an attempt to minimize image mosiac seamlines, a Cubist machine learning approach to radiometric normalization has been implemented to match radiometry from imagery taken at different dates and by differenet sensors.  In this study, Satellogic 4-band Mark V MS imagery was co-registered to Maxar Worlview-03 Imagery and a Cubist machine learning workflow was devleoped to match the Sateloggic TOA pixel values to the Surface Reflectance values of the Maxar imager.  The Maxar and Satellogic data were combined to create a 4-band image mosaic with seamless boundaries and a spectral integrity necessary to derive analytical products such as land cover and change detection.  This approach of fusing different multi-spectral sensors will allow for the use of a substantially larger data archive to pull from when it comes to generating remote sensing products, including greater geospatial coverage and higher temporal frequency.
Eric Jeronimus, Maxar