Abstracts Track 2024

Area 1 - GIS and Climate Change

Nr: 54

Insights into Air Temperature Variability Among Local Climate Zones in the Metropolitan City of Milan


Afshin Moazzam, Alberto Vavassori, Maria A. Brovelli, Daniele Oxoli, Giovanna Venuti, Mario Siciliani de Cumis, Patrizia Sacco and Deodato Tapete

Abstract: Effective policy formulation to address the urgent global challenges of climate change necessitates the establishment of robust monitoring systems coupled with constant data analysis and management. In this context, the LCZ-ODZ project (agreement n.2022-30-HH.0), funded by the Italian Space Agency (ASI), is demonstrating an innovative analytical framework to generate Local Climate Zone (LCZ) maps. The demonstration is carried out over the Metropolitan City of Milan, Italy by integrating hyperspectral PRISMA and multispectral Sentinel-2 satellite images and evaluating the correlation between LCZ maps and air temperature measurements across the study region. The primary focus of this abstract is on reporting the findings on the LCZ and air temperature correlation. In our research, we have leveraged three distinct data sources. Firstly, the Lombardy Region Environmental Protection Agency (ARPA Lombardia) meteorological network providing authoritative ground-sensed temperature measurements with 10-minute temporal frequency. The temperature measurements are gathered using The ARPA Weather Sensors Plugin which has been developed within the LCZ-ODC project to download, process, and plot the time series data directly with QGIS. Secondly, interpolated air temperature maps from the ClimaMi project depicting the spatial distribution of average and extreme air temperatures near the ground's atmospheric layer on a standard 100m×100m grid. Lastly, the Visual Crossing global weather database encompassing weather information and synoptic data from worldwide observation stations. 12 ARPA stations actively collecting temperature measurements within the study region have been selected for the analysis. Based on the LCZ maps produced within the project, each station has been assigned one LCZ class. However, considering adjacent classes often exhibit a continuous thermal regime, we computed the fractions of LCZ classes falling within the 200m buffer around each ARPA station and we assigned the LCZ type covering the majority (> 55%) of the buffer area. A subclass (i.e., LCZ type covering > 20% of the buffer) was also assigned to each station to comprehensively describe varying structures and land cover inside the buffer. Comparing the temporal variation of LCZ maps from February to August 2023 obtained from both satellites, it appeared that a single LCZ type could be assigned to each station. To continue the analysis, we initially removed outliers from the ARPA temperature data using the interquartile range (IQR) method, given its non-normal distribution. We considered temperature data between 2 p.m. and 4 p.m. as daytime measurements, where peak temperatures typically occur, and between 3 to 5 hours after sunset for night-time measurements. Accordingly, day and night temperature differences across the different LCZs were investigated for February (Winter) and August (Summer) 2023. Our findings indicate that natural LCZs show greater temperature variations between day and night compared to built-up areas. Built-up urban zones exhibit higher nighttime temperatures, forming urban heat islands (UHI). Daytime temperature differences among classes are minimal (< 1°C), suggesting consistent thermal behavior. Seasonal differences reveal more pronounced temperature variations between day and night during summer than winter across all LCZs. Future analyses will consider the effects of cloud cover and wind speed on the results.

Area 2 - Interaction with Spatial-Temporal Information

Nr: 56

Development and Application of Sentinel-2 Satellite Imagery Datasets for Deep-Learning Driven Forest Change Detection


Valeria Martin Hernandez

Abstract: Forest loss due to natural events, such as wildfires, represents an increasing global challenge that demands advanced analytical methods for effective detection and mitigation. To this end, the integration of satellite imagery with deep learning (DL) methods has become essential. Nevertheless, this approach requires substantial amounts of labeled data to produce accurate results. In this study, we use bi-temporal Sentinel-2 satellite imagery sourced from Google Earth Engine (GEE) to create the California Wildfire GeoImaging Dataset (CWGID), a high-resolution labeled satellite imagery dataset with over 100,000 labeled image pairs for wildfire detection through DL applications. Our methods include data acquisition from authoritative sources, data processing, and an initial analysis of the dataset using Convolutional Neural Network (CNN) architectures. The results show that the EfficientNet-B0 model achieves the highest accuracy of over 92% in detecting forest wildfires. The CWGID and the methodology used to build it, prove to be a valuable resource in the training needed by DL architectures for accurate forest wildfire detection. Future studies will aim to extend this research to other types of forest loss, such as destructive windstorms and droughts.

Nr: 34

Utilizing Satellite Remote Sensing for an Initial Forensic Inquiry into Elevated Internal Temperatures in Landfills


Rouzbeh Nazari and Maryam Karimi

Abstract: Subsurface fires and smoldering events in landfills present substantial health and environmental risks, with extinguishing efforts proving more challenging and costly than open surface fires. The initiation of subsurface fires often remains unnoticed for extended periods, allowing them to spread extensively. Compounding the issue, not all landfills maintain or disclose heat elevation data, and some lack systems for extracting gases to regulate subsurface temperatures. The prompt and cost-effective detection of subsurface fires emerges as a critical concern. In this study, we propose a method utilizing moderate spatial resolution satellite thermal infrared imagery to pinpoint the location of subsurface fires and monitor their migration within landfills. Our focus is on the Bridgeton Sanitary Landfill in Bridgeton, MO, where a subsurface fire was initially identified in 2010 and persists to this day. Over the past seventeen years, Landsat satellite observations have been scrutinized for surface temperature anomalies, or hot spots, potentially linked to subsurface fires. Our findings reveal a significant correlation between the identified hot spots in satellite imagery and the documented locations of subsurface fires. The study further explores changes in hot spot locations over time, aligning with the spreading routes of subsurface fires determined through in situ measurements. This research demonstrates the efficacy of our proposed satellite-based approach as a valuable tool for identifying subsurface fires in landfills. The results emphasize the practical application of satellite observations for landfill owners and operators, offering a means to monitor landfills effectively and reduce expenses associated with extinguishing subsurface fires. The proposed method not only enhances the timely identification of subsurface fires but also contributes to proactive measures in minimizing the environmental and financial impacts of landfill fires.

Area 3 - Managing Spatial Data

Nr: 55

Size and Shape LiDAR Footprint: A Necessary Analysis to Understand a Positional Uncertainty


Juan F. Reinoso Gordo, Francisco Ariza-López, Juan J. González-Quiñones and Manuel Ureña-Cámara

Abstract: Amongst the most important factors influencing the positional uncertainty (x, y, z coordinates) of any point in the point cloud generated by a LiDAR is the individual footprint resulting from the intersection of a laser beam on the ground. Due to the laser divergence, the beam is not transmitted as a straight line but as a cone and the projection on the terrain of this beam would not be a point but a conical surface, which in case of considering the terrain as a plane, this intersection would be an ellipse. For this reason, the value of the point coordinates (x, y, z) on the terrain would actually be affected by the coordinates of all the points on the terrain that fall within the mentioned ellipse, that is inside the footprint. Therefore, to estimate the positional uncertainty of every point in the point cloud it will be necessary a previous study of this factor as a dependent variable of the following independent variables: laser divergence (αlpha, expressed as an angle), aircraft altitude (H, LiDAR-terrain distance), and the angle of incidence on the terrain (γ, dependent on the terrain slope and orientation with respect to the laser). To perform a first analysis of the footprint it is convenient to simplify the terrain shape to a plane, so that the cone intersection on a flat surface allows to compute the semi-major (a) and minor (b) axis from the ellipse, as well as the flattening parameter. There is an additional data that is of interest in this study and that is the maximum distance (md) between the point M (intersection of the axis of the cone with the plane) and the point on the ellipse farthest from that point M. That md is greater than the semi-major axis of the ellipse since the axis of the cone, although it intersects with the plane on the major diameter of the ellipse does not do it in the center of the same one (it only intersects in the center of the ellipse when the angle of incidence is 90º, that is, when the ellipse is a circumference). Also of interest is the mean radius of the ellipse defined as the integral rm=1/2π ∫_0^2π〖√((a cosθ )^2+(a sinθ )^2 ) dθ〗 and the variance of the radius with respect to rm defined as 〖Var〗_rm=1/2π ∫_0^2π〖(rm- √((a cosθ )^2+(a sinθ )^2 ))^2 dθ〗. Both expressions have to be solved by numerical integration. In a first approach to the study of a LiDAR footprint, the Leica ALS50-II, whose divergence is 0.00022 radians, has been selected and the evolution of the variables for 4 aircraft heights of 500, 1000, 1500 and 2000 m have been calculated. Figure 1 represents the plots relative to a and b semiaxis. From the study of the values of a, b, dm, rm and Varrm it has been observed that: 1) the differences between a and dm are practically negligible because the divergence value is very small and it is to be expected for most LiDAR devices, 2) the variability of b along the different angles of incidence for the same altitude is small (horizontal trend in figure 1), and 3) the Varrm variability is significant both if analyzed from the point of view of the different heights, and if analyzed taking into account only the angle of incidence. For this reason, the standard deviation derived from the Varrm seems to be a candidate value to be used as a scaling factor for positional uncertainty computation. Acknowledgements: Grant PID2022-138835NB-I00 funded by the Spanish MICIU/AEI/10.13039/501100011033 and by “ERDF/EU".

Area 4 - Remote Sensing

Nr: 13

High Resolution Remote Sensing for Biodiversity Assessment: A Case Study of Tree Species in an Old Growth Forest


Yousef Erfanifard, Krzysztof Stereńczak and Maciej Lisiewicz

Abstract: Traditional field measurements (TFMs) for assessing the diversity of tree species in forests, while valuable, are inherently labor-intensive, time-consuming, and often impractical for vast and remote forest areas. In light of these challenges, the integration of high resolution remote sensing datasets, such as airborne laser scanning (ALS) point clouds and color infrared (CIR) images, emerges as a promising solution to overcome the limitations of TFMs. In the present study, the biodiversity indices were initially computed at plot levels based on TFMs and recognized individuals on the ALS and CIR datasets separately. Finally, a comparative analysis was conducted between the results obtained from our approach and those derived from TFMs. The study is conducted in the Bialowieza Forest (BF), a renowned UNESCO world heritage site recognized as an old growth forest. For effective management, the forest is categorized into three distinct parts: strict reserve (SR), nature reserve (NR), and managed forest (MF). A total of 570 sample plots, each covering an area of 500 m2 and previously surveyed for species, diameter at breast height (dbh), and height, were chosen in SR (168 plots), NR (199 plots) and MF (203 plots). Utilizing ALS and CIR datasets, 30 variables were extracted and employed as input for the RF algorithm to identify 12 genera of broadleaved and coniferous individuals within the BF in the previous study. We computed the biodiversity indices, including alpha diversity, beta diversity, gamma diversity, Simpson, Shannon-Wiener, density, dominance, and importance, using measurements derived from both TFMs and ALS+CIR datasets for the individuals that were observable from above (i.e., dominant and intermediate individuals). The findings revealed no significant differences in alpha diversity indices (Shannon-Wiener and Simpson) between TFMs (SR: 0.50, 0.32; NR: 1.17, 0.66; MF: 0.65, 0.36, respectively) and ALS+CIR datasets (SR: 1.27, 0.67; NR: 1.09, 0.66; MF: 1.19, 0.62, respective) across the three management types (p-value = 0.2435, 0.1828). Average beta diversity, computed from TFMs (0.69) and ALS+CIR (0.72) datasets, exhibited no statistically significant difference. Similarly, Wilcoxon rank test results indicated no distinction (p-value = 0.9047) between average gamma diversities derived from TFMs (2.10) and ALS+CIR (2.16) datasets. Notably, the top two species in terms of density, dominance, and importance were accurately identified using ALS+CIR datasets. In addition, principal component analysis (PCA) illustrated a robust correlation (~0.94) between PCA1 and PCA2 across both TFMs and ALS+CIR datasets, highlighting the consistency of diversity patterns among the various forest management types. The comparison of diversity indices between TFMs and ALS+CIR datasets across the distinct management types suggests that the high-resolution remote sensing approach provided comparable results to traditional field methods in capturing the diversity of tree species within these ecosystems. In conclusion, this study demonstrates the viability of utilizing ALS+CIR datasets for biodiversity assessment in the BF across various management types. Overall, the integration of ALS and CIR datasets emerges as a promising avenue for advancing our understanding of forest biodiversity and facilitating evidence-based decision-making for the conservation of biodiversity in old growth forests.

Nr: 20

Continental-Scale Mapping of Banana for BBTV Mitigation Using Remote Sensing and AI Tools


Tunrayo R. Alabi, Patrick O. Duke and Lava Kumar

Abstract: Banana and plantain are commonly grown as mixed crops by smallholder farmers in backyards and small farmlands in sub-Saharan Africa (SSA). Unfortunately, the crop is at risk from various pests and diseases, including the banana bunchy top virus (BBTV), which is invasive and poses a significant threat to banana production. BBTV is spreading to new areas across Africa, having originated in the Benin Republic in 2010 and since spreading to Nigeria (2011) and Togo (2018). Surveillance to assess the disease in banana fields is difficult due to small and fragmented production spread across large areas, which makes conventional field survey methods time-consuming and expensive. To tackle this challenge, we have developed remote sensing-based Artificial Intelligence (AI) tools to identify banana fields for targeted BBTV surveillance. We used synthetic aperture radar (SAR) data from the Copernicus Sentinel-1 of the European Space Agency(ESA) satellite and AI algorithms such as random forest (RF) and convolutional neural networks (CNN) in the banana-growing areas of Nigeria, Benin, Togo and Ghana. We used Unmanned Aerial Vehicle (UAV) RGB images and the GoPro action camera photos to train and validate the models. Our results show that we can accurately identify and monitor banana fields at an accuracy of about 80%, demonstrating great potential to positively impact the health and sustainability of banana production in African smallholder systems. Additionally, our findings indicate that the remote sensing-based AI models for mapping bananas can provide crucial input for designing surveillance programs for BBTV. It can improve precision and efficiency for guided surveillance for detecting and mapping the banana bunchy top virus (BBTV) spread and supports data-informed decision-making on BBTV containment strategies in sub-Saharan Africa. Supplementary material Alabi, T. R., Adewopo, J., Duke, O. P., & Kumar, P. L. (2022). Banana Mapping in Heterogenous Smallholder Farming Systems Using High-Resolution Remote Sensing Imagery and Machine Learning Models with Implications for Banana Bunchy Top Disease Surveillance. Remote Sensing, 14(20), 5206. https://doi.org/10.3390/rs14205206.

Nr: 30

Use of Multispectral UAV-Drone and Sentinel-2 Imagery for the Prediction of Dissolved Oxygen in Dam Reservoir with Stacked Ensemble Machine Learning Models


Phillimon T. Odirile, Yashon O. Ouma, Boipuso Nkwae, Ditiro B. Moalafhi, George Anderson, Bhagabat P. Parida and Jiaguo Qi

Abstract: Timely and continuous monitoring of reservoir water quality is critical for sustainable management of the water resource. In most developing countries however, water management authorities rely on in-situ water quality monitoring, which is costly, time-consuming and do not cover the entire water body. Equipped with spectral sensors, the use of UAV-drones and satellite sensors can compensate for these limitations by cost-effectively acquiring high spatiotemporal resolution data for water quality monitoring. This study utilized multispectral data from DJI-P4 UAV-drone and Sentinel-2 MSI satellite data for the retrieval of dissolved oxygen (DO) in Gaborone Dam (Botswana) using machine learning (ML) models. DO is one of the most vital water quality parameters (WQPs) that directly indicates the health of the water body and the survival of aquatic life. The in-situ measurements comprised of 21 water samples (15 for training, 6 for testing) collected in November 2022. For most of the optically inactive WQPs such as DO, the complex bio-optical environment with different pollutants makes it difficult to determine the specific internal and nonlinear correlations between the spectral information and the optically inactive WQP. Though individual ML algorithms have demonstrated significant improvements in the inversion of optically inactive WQPs, due to the ability to resolve underfitting problems, they tend to be prone to overfitting and exhibit poor generalization capabilities. This study compared five ensemble ML models, Random Forest (RF), eXtreme Gradient Boosting (XGBoost), Adaptive Boosting (AdaBoost), CATegorical Boosting (CATBoost), and ExtraTrees for the inversion of DO from the sensors. To minimize overfitting and improve on the DO retrieval accuracy, this study proposes stacked ensemble ML (SE-ML) approach that combines the two best MLs. The input ML prediction datasets include 84 spectral indices from visible (Blue, Green, Red), Near-Infrared Red (NIR), and Red-Edge bands. From the UAV-drone data, XGBoost (R2 = NSE = 0.926, MAE = 0.125 mg/L, RMSE = 0.407 mg/L, PBIAS = -9.564) and CATBoost (R2 = NSE = 0.914, MAE = 0.181 mg/L, RMSE = 1.696 mg/L, PBIAS = -8.031) outperformed RF, AdaBoost, and ExtraTrees in the retrieval of DO. For the UAV-drone data, the Stacked XGBoost and CATBoost improved the DO retrieval accuracy to R2 = NSE = 0.957; MAE = 0.099 mg/L; RMSE = 0.142 mg/L and PBIAS = 0.079. Using Sentinel-2 data, XGBoost and CATBoost also performed better than the other ML algorithms with similar accuracy results, R2 = NSE = 0.909; MAE = 0.578 mg/L; RMSE = 1.741 mg/L, and PBIAS <-0.1. The proposed SE-ML enhanced the estimation of DO from Sentinel-2 with R2 = NSE = 0.970; MAE = 0.077 mg/L; RMSE = 0.011 mg/L and PBIAS = 0.184. Examination of the spectral feature combinations identified NIR band as most significant for DO estimation from UAV-drone, while Red and NIR wavelengths contributed most to DO prediction from Sentinel-2. With the potential of providing high-frequency and large spatial coverage observational data in near-real-time, this study demonstrates the potential of UAV-drone and stacked ensemble ML algorithms for the prediction of dissolved oxygen in dam water reservoirs. Ongoing and further investigations are focusing on the replicability and applicability of the SE-ML across different temporal resolutions and offering broader insights in water quality monitoring using advanced sensing and modelling techniques.

Area 5 - Domain Applications

Nr: 53

Geospatial Analysis of Health, Environmental Justice and Noise Exposure in Florida, USA


Yelena Ogneva-Himmelberger

Abstract: Objective In this paper, we use GIS to identify areas affected by high noise from aviation, roads and railroad traffic in Miami-Dade county in Florida, USA and to analyze the prevalence of chronic disease in these areas. Specifically, we address three questions: How many people are exposed to high noise level? Are there racial/ethnic inequities in exposure to noise? Is noise exposure significantly higher in hot spots than in cold spots of the four health-related measures (high blood pressure, coronary heart disease, stroke, and lack of sleep)? High noise exposure a risk factor for these chronic diseases, so it is important to know where both of these conditions are co-occurring. Our findings could inform the development of effective programs and policies related to noise abatement, and help prioritize investment in areas with the biggest health inequities. Data Aviation noise raster (30-meter) was obtained from the Bureau of Transportation Statistics. It represents the average daily noise level, modeled using flights, meteorological data, and road and railroad traffic data. Health outcomes data was obtained from PLACES database (Population Level Analysis and Community Estimates) at census tract level. PLACES database provides local-level model-based estimates for 29 health measures. We selected 4 measures known to be associated with chronic noise exposure – the prevalence of high blood pressure, coronary heart disease, stroke, and sleep deprivation. Methods We reclassified noise data into “medium noise” (45-65 db) and “high noise” (65db+) categories, and overlaid them with detailed population data (census block) to calculate total population, Black, Hispanic, and White population inside and outside noisy areas. The Federal Aviation Administration uses 65db as the threshold for “significant” noise exposure. Hot Spot Analysis (Getis-Ord Gi*) technique was applied to four health measures, and census tracts within hot and cold spots (>95% confidence) were selected. Zonal statistic tool was used with the noise raster to calculate average noise level for each census tract. Independent samples T-test was used to compare noise exposure inside hot and cold spots of health measures. Results About 48% of the county population (1,303,141) lives in areas with daily noise level over 45dB, including about 35,000 persons in areas of significant noise exposure (>65 dB). There are no racial/ethnic inequities in exposure to noise because population proportions are similar in low, medium, and high noise areas. The mean values for noise were statistically significantly higher in hot spots of three health measures compared to cold spots (high blood pressure, coronary heart disease, and stroke). Conclusion We analyzed two novel, high resolution, nationwide data sets and found a statistically significant association between spatial clusters of prevalence of three chronic diseases and noise levels. Hot spots are areas where a high proportion of the population is estimated to have poor health condition, and is simultaneously experiencing chronic exposure to very high noise levels from road, railroads and airports. These findings can support practitioners and local policy makers in their efforts to improve the conditions of vulnerable populations in these areas.