TS-SatFire: A Multi-Task Satellite Image Time-Series Dataset for Wildfire Detection and Prediction.

Due to the irregular and sporadic nature of wildfires, continuous monitoring of large areas is required. Since geostationary satellite sensors can observe large areas with high temporal resolution, they are suitable for monitoring wildfires in real time. However, the threshold algorithm currently employed for satellite-based active fire detection has poor performance in sensors with low spatial resolution. In addition, the algorithm does not account for environmental conditions that affect wildfire detection, resulting in poor generalization performance for large areas. This study examines the viability of an adaptive active fire detection model by combining satellite and numerical model data with deep learning. A model for active fire detection was developed using commonly employed brightness temperature-related variables (key variables) and local environmental variables (sub variables). Key variables are the cross spectral and spatial differences between the MIR (central wavelength of 3.85 m) and 2 TIR (central wavelengths of 9.63 and 11.20 m) channels of the Advanced Himawari Imager (AHI). Sub variables include Solar zenith angle (SOZ) and satellite zenith angle (SAZ) of AHI, skin temperature (ST) and relative humidity (RH) of European Centre for Medium-Range Weather Forecasts (ECMWF) Reanalysis v5 (ERA5)-land data. Four processes (confidence, frequency, land cover, and continuity tests) were used to extract reference fire samples from Moderate Resolution Imaging Spectroradiometer (MODIS) and Visible Infrared Imaging Radiometer Suite (VIIRS) active fire products. To consider the different properties of key and sub variables, a 2-way convolutional neural network (CNN) structure was developed. To evaluate the influence of environmental variables, a CNN model without sub variables was adopted as a control model. The 2-way CNN (recall of 0.86, precision of 0.96, and standard deviation of recall of 0.13) was more robust at five focus sites than the control CNN (recall of 0.82, precision of 0.97, and standard deviation of recall of 0.163). Despite having a lower spatial resolution than MODIS/VIIRS, 2-way CNN outperformed other satellite-based active fire products (MODIS, VIIRS, AHI, and Advanced Meteorological Imager) in terms of detection capacity. The control CNN demonstrated poor performance under certain environmental conditions (high RH, high SAZ, and transition time between day and night), but 2-way CNN mitigates this tendency. In particular, the use of RH improved detection sensitivity, and SAZ contributed to the spatial robustness. This study demonstrated the significance of environmental conditions in active fire detection and proposed a suitable CNN structure for this intent. Based on the findings of this study, higher-level adaptive active fire monitoring under diverse environmental conditions will be possible together with explainable artificial intelligence.

Fire in Brazilian Pantanal represents a serious threat to biodiversity. The Brazilian National Institute of Spatial Research (INPE) has a program named Queimadas, which estimated from January 2020 to October 2020, a burned area in Pantanal of approximately 40,606 km2. This program also provides daily data of active fire (fires spots) from a methodology that uses MODIS (Aqua and Terra) sensor data as reference satellites, which presents limitations mainly when dealing with small active fires. Remote sensing researches on active fire dynamics have contributed to wildfire comprehension, despite generally applying low spatial resolution data. Convolutional Neural Networks (CNN) associated with high- and medium-resolution remote sensing data may provide a complementary strategy to small active fire detection. We propose an approach based on object detection methods to map active fire in the Pantanal. In this approach, a post-processing strategy is adopted based on Non-Max Suppression (NMS) to reduce the number of highly overlapped detections. Extensive experiments were conducted, generating 150 models, as five-folds were considered. We generate a public dataset with 775-RGB image patches from the Wide Field Imager (WFI) sensor onboard the China Brazil Earth Resources Satellite (CBERS) 4A. The patches resulted from 49 images acquired from May to August 2020 and present a spatial and temporal resolutions of 55 m and five days, respectively. The proposed approach uses a point (active fire) to generate squared bounding boxes. Our findings indicate that accurate results were achieved, even considering recent images from 2021, showing the generalization capability of our models to complement other researches and wildfire databases such as the current program Queimadas in detecting active fire in this complex environment. The approach may be extended and evaluated in other environmental conditions worldwide where active fire detection is still a required information in fire fighting and rescue initiatives.

The array of wildfire activities instigated by human endeavors has emerged as a significant source of atmospheric pollution, posing considerable risks to both public health and property safety. This study harnesses Sentinel-2 satellite data, employing a variety of methods including spectral index methods, thresholding, and the Random Forest (RF) model for active fire spot detection. The research encompasses a wide range of land cover types across various Chinese regions. Utilizing the Gini coefficient, the study assesses the importance of spectral and texture features in the RF, culminating in the selection of an optimal feature combination for the construction of a bespoke RF model tailored for active fire detection. The research utilized texture features based on the Grey Level Co-occurrence Matrix (GLCM), demonstrating their significant contribution to enhancing the accuracy of fire detection using the RF model. Our analysis reveals that GLCM-based texture features, which form 40% of the model’s final feature set, are crucial for improving detection accuracy. The optimized RF model demonstrates a marked superiority in identifying active fires, achieving an overall accuracy of 86.1%. The study results demonstrate that the bespoke RF model is suitable for detecting active fire across various land cover environments in China.

In contrast with current operational products of burned area, which are generally available one month after the fire, active fires are readily available, with potential application for early evaluation of approximate fire perimeters to support fire management decision making in near real time. While previous coarse-scale studies have focused on relating the number of active fires to a burned area, some local-scale studies have proposed the spatial aggregation of active fires to directly obtain early estimate perimeters from active fires. Nevertheless, further analysis of this latter technique, including the definition of aggregation distance and large-scale testing, is still required. There is a need for studies that evaluate the potential of active fire aggregation for rapid initial fire perimeter delineation, particularly taking advantage of the improved spatial resolution of the Visible Infrared Imaging Radiometer (VIIRS) 375 m, over large areas and long periods of study. The current study tested the use of convex hull algorithms for deriving coarse-scale perimeters from Moderate Resolution Imaging Spectroradiometer (MODIS) and Visible Infrared Imaging Radiometer Suite (VIIRS) active fire detections, compared against the mapped perimeter of the MODIS collection 6 (MCD64A1) burned area. We analyzed the effect of aggregation distance (750, 1000, 1125 and 1500 m) on the relationships of active fire perimeters with MCD64A1, for both individual fire perimeter prediction and total burned area estimation, for the period 2012–2108 in Mexico. The aggregation of active fire detections from MODIS and VIIRS demonstrated a potential to offer coarse-scale early estimates of the perimeters of large fires, which can be available to support fire monitoring and management in near real time. Total burned area predicted from aggregated active fires followed the same temporal behavior as the standard MCD64A1 burned area, with potential to also account for the role of smaller fires detected by the thermal anomalies. The proposed methodology, based on easily available algorithms of point aggregation, is susceptible to be utilized both for near real-time and historical fire perimeter evaluation elsewhere. Future studies might test active fires aggregation between regions or biomes with contrasting fuel characteristics and human activity patterns against medium resolution (e.g., Landsat and Sentinel) fire perimeters. Furthermore, coarse-scale active fire perimeters might be utilized to locate areas where such higher-resolution imagery can be downloaded to improve the evaluation of fire extent and impact.

In this paper, the optimal threshold for fire detection is derived based on the probability density functions of the fire and background pixels. The threshold values and the commission and omission errors are computed according to several detection criteria: maximum likelihood, null-hypothesis testing, and minimum cost. This statistical approach enables the determination of optimal threshold value according to the specific requirements of the detection tasks I. INTRODUCTION Active fire detection using satellite thermal sensors usually involves thresholding the detected brightness temperature in several bands. The frequently used features for fire detection are the brightness temperature in the 4-µm wavelength band (T4) and the brightness temperature difference between the 4- µm and 11- µm bands (∆T=T4-T11) (1), (2), (3). Due to the statistical nature of the fire and background temperature distribution, no matter how the thresholds are chosen, there are bound to be commission errors (i.e. false alarms) and omission errors (false negatives) in the detection outcomes. In the commonly used fire detection algorithms, the thresholds are often arbitrarily defined, and hence may not be optimal for a given fire detection task. The optimal threshold values are expected to be site dependent. They are different in different regions and for different fire conditions (such as smoldering and flaming fires). Hence, a global set of threshold values may not work optimally for a specific site. The task of active fire hot spot detection can be modeled as a stochastic target detection procedure (4). In this model, each pixel of the image occupies a point (T4, ∆T) in the two- dimensional feature space, and each pixel belongs to either the fire class or the non-fire background class. This approach considers the probability density functions of the fire and background pixels and optimal thresholds are derived depending on the specific objectives of the detection tasks. In a previous work (4), we implemented a method for deriving the optimal detection threshold by minimizing a cost function which is a weighted sum of the omission and commission errors. Alternatively, the threshold can also be derived based on other criteria. In this paper, we derive the optimal thresholds for detecting active fires in MODIS scenes using various detection criteria: maximum likelihood, constant false alarm rate, and null-hypothesis testing. The test data set consists of several MODIS scenes over Sumatra and Borneo islands and coincidental high resolution SPOT scenes (5). The SPOT data serves as ground truth for validation of MODIS hotspots detected by the new algorithms based on stochastic models. The probability density functions (pdf's) of the fire and background pixels are derived from this dataset and these pdf's are used to derive the optimal thresholds for active fire detection.

Multi-resolution monitoring of the 2023 maui wildfires, implications and needs for satellite-based wildfire disaster monitoring

Remote Sensing of Biomass Burning in Tropical Regions: Sampling Issues and Multisensor Approach

The Sea and Land Surface Temperature Radiometer (SLSTR) senses the Earth from onboard two concurrently operating European Copernicus Sentinel-3 (S3) satellites. As the Terra platform carrying the Moderate Resolution Imaging Spectroradiometer (MODIS) is reaching its end of life, the S3 Active Fire Detection and FRP products generated from data captured by S3 SLSTR are expected to soon become the main global active fire (AF) product for the mid-morning and evening low Earth orbit timeslots. The S3 night-time AF product issued by the European Space Agency (ESA) has been operational since March 2020, and here we report on the significant adjustments made to enable the generation of a complimentary daytime product. Similar to MODIS, SLSTR possesses two middle infrared channels, both a ‘standard’ (normal gain; S7) channel and a ‘fire’ (low-gain; F1) channel - but in contrast to MODIS by day even the ambient background land surface is often saturated in the SLSTR standard gain MIR (S7) channel. This saturation necessitates far greater use of the F1 channel data by day for active fire detection than by night, even though F1 has characteristics which make its data more challenging to combine with that from the other SLSTR thermal infrared channels. Here we report on the approaches used to combine S7 and F1 data for optimized daytime AF detection, and also detail the other algorithm adjustments found necessary to include in the daytime AF product algorithm. We compare the resulting daytime SLSTR AF product data to that generated from near-simultaneous views provided by MODIS onboard Terra. When both sensors detect the same active fire cluster at similar time, there is minimal bias shown between the two FRP retrievals (the ordinary least squares linear best fit between matched SLSTR and MODIS per-fire FRP matchups has a slope of 0.97). At the regional scale, the S3 product detects 70% of the AF pixels that the matching MODIS product reports, but also provides a further (16%) set of unique AF pixel detections. Regional FRP totals derived from SLSTR appear slightly lower than those from MODIS, and the OLS linear best fit between these regional FRP matchup datasets has a slope of 0.91. This is largely due to SLSTR performing less well in detecting the lowest FRP fires by day, whereas by night the S3 product performs a little better than MODIS due to the increased night-time use of S7 in the earlier AF pixel detection stages. Global fire mapping at a 0.25° grid cell resolution shows very similar daytime fire patterns and FRP totals from S3 and Terra MODIS, with SLSTR detecting around twice the number of AF pixels due to the algorithm being more effective at identifying low FRP pixels at the edges of fire clusters. Regional time series case studies also show very similar temporal patterns between S3 and Terra MODIS. Longer-term intercomparisons such as these will provide the knowledge necessary to use MODIS and SLSTR AF products together to analyse long-term AF trends. Comparing near simultaneous observations of fires by SLSTR and from the 30 m spatial resolution Landsat Operational Land Image (OLI) data, we find that once there are around 150 OLI active fire pixels detected within the area of an SLSTR pixel, the chances of that SLSTR pixel being classed as an active fire by the daytime algorithm rises to almost 100%. The daytime SLSTR AF Detection and FRP product based on the algorithm described herein has been fully operational since March 2022 and is available from the Sentinel-3 Science Hub (https://scihub.copernicus.eu/).

To develop and externally validate a computed tomography (CT)-based multitask learning model to predict fracture risk. This study was conducted in two parts, using a multitasking learning approach. We developed a cross-sectional vertebral fracture (VF) detection model using abdominal CT scans of 2553 patients aged 50-80 years. Then, we leveraged this detection model within a multitask learning framework to develop a longitudinal VF prediction model over a 5-year follow-up period. External testing was performed on 1506 patients from two independent hospitals. The performance was compared between the single-task and multitask models, bone-only and bone+muscle images, and image-only and clinical models. For the cross-sectional fracture detection model, the mean age of the patients was 76.2 years, and 66.7% were female. In the classification task for detection of VF, the model using both bone and muscle showed an area under the receiver operating characteristic curve (AUROC) of 0.82 in the development set and 0.80 in the external test sets. Using multitask learning, the bone + muscle image model showed a c-index of 0.68 and had superior performance than the bone-only model in the external test set for 2-year, 3-year, and 5-year AUROCs (0.79 vs. 0.75, 0.71 vs. 0.68, and 0.71 vs. 0.68, respectively, all p < 0.01). Also, the multitask model significantly outperformed the Fracture Risk Assessment Tool (FRAX) (c-index: 0.68 vs. 0.66, p < 0.01). The CT-based multitask learning model integrating both bone and muscle data showed superior predictive performance for VFs compared with models using bone images only and traditional clinical models. Question Vertebral fracture risk remains underestimated in many individuals undergoing CT scans for other reasons, highlighting the need for improved opportunistic prediction tools. Findings A multitask deep learning model integrating both bone and muscle features from CT scans demonstrated superior performance compared to bone-only and traditional clinical models, including FRAX. Clinical relevance The proposed model enables accurate vertebral fracture risk prediction using routinely acquired CT scans, facilitating early identification and intervention without the need for additional tests.

The Upper Guinean region of West Africa exhibits strong geographic variation in land use, climate, vegetation, and human population and has experienced phenomenal biophysical and socio-economic changes in recent decades. All of these factors influence spatial heterogeneity and temporal trends in fires, but their combined effects on fire regimes are not well understood. The main objectives of this study were to characterize the spatial patterns and interrelationships of multiple fire regime components, identify recent trends in fire activity, and explore the relative influences of climate, topography, vegetation type, and human activity on fire regimes. Fire regime components, including active fire density, burned area, fire season length, and fire radiative power, were characterized using MODIS fire products from 2003 to 2015. Both active fire and burned area were most strongly associated with vegetation type, whereas fire season length was most strongly influenced by climate and topography variables, and fire radiative power was most strongly influenced by climate. These associations resulted in a gradient of increasing fire activity from forested coastal regions to the savanna-dominated interior, as well as large variations in burned area and fire season length within the savanna regions and high fire radiative power in the westernmost coastal regions. There were increasing trends in active fire detections in parts of the Western Guinean Lowland Forests ecoregion and decreasing trends in both active fire detections and burned area in savanna-dominated ecoregions. These results portend that ongoing regional landscape and socio-economic changes along with climate change will lead to further changes in the fire regimes in West Africa. Efforts to project future fire regimes and develop regional strategies for adaptation will need to encompass multiple components of the fire regime and consider multiple drivers, including land use as well as climate.

We demonstrate a new active fire (AF) detection and characterisation approach for use with the VIIRS spaceborne sensor. This includes for the first-time joint exploitation of both 375m I-Band and 750m M-Band data to provide both AF detections and FRP (fire radiative power) retrievals over the full range of fire and FRP magnitudes. We demonstrate the value of our VIIRS-IM ‘synergy’ product in an area of eastern China dominated by numerous small agricultural residue burns, which contribute significantly to regional air quality problems but which are often difficult to identify via standard (e.g. MODIS 500m resolution) burned area mapping. We show that the highly ‘fire sensitive’ VIIRS I-Band data enables detection of the ‘small’ active fires (FRP≤1MW), but this sensitivity can lead to false alarms, often associated with manmade structures. We help avoid these via use of 30m resolution global land cover data and an OpenStreetMap mask. Comparisons to near-simultaneous Aqua-MODIS AF detections, and the existing VIIRS I-Band AF global product, highlight our VIIRS algorithm's ability to more reliably detect the lowest FRP pixels, associated with the type of agricultural burning dominating eastern China. Our algorithm delivers typically 5 to 10× more AF pixels than does simultaneous-collected MODIS AF data (notwithstanding differences in spatial resolution), and importantly with a AF detection sensitivity that remains much more constant across the swath due to VIIRS' unique pixel aggregation scheme. The VIIRS I4-Band saturates over higher FRP fires, but by combining use of I- and M-Band data our algorithm generates reliable FRP records for all fires regardless of FRP magnitude. Using the VIIRS-IM methodology we find regionally summed FRP's up to 4× higher than are recorded by MODIS over the same fire season, highlighting the significance of the formally undetected low FRP active fires and indicating that current MODIS FRP-based emissions inventories for areas dominated by agricultural burning may be underestimating in a similar way to burned-area based approaches. FRP generation from VIIRS that takes into account both low- and high-FRP fires via use of both the I- and M-Band data should therefore enable significant improvements in global fire emissions estimation, particularly for regions where smaller types of fire are especially dominant.

BackgroundGenomic prediction in multiple populations can be viewed as a multi-task learning problem where tasks are to derive prediction equations for each population and multi-task learning property can be improved by sharing information across populations. The goal of this study was to develop a multi-task Bayesian learning model for multi-population genomic prediction with a strategy to effectively share information across populations. Simulation studies and real data from Holstein and Ayrshire dairy breeds with phenotypes on five milk production traits were used to evaluate the proposed multi-task Bayesian learning model and compare with a single-task model and a simple data pooling method.ResultsA multi-task Bayesian learning model was proposed for multi-population genomic prediction. Information was shared across populations through a common set of latent indicator variables while SNP effects were allowed to vary in different populations. Both simulation studies and real data analysis showed the effectiveness of the multi-task model in improving genomic prediction accuracy for the smaller Ayshire breed. Simulation studies suggested that the multi-task model was most effective when the number of QTL was small (n = 20), with an increase of accuracy by up to 0.09 when QTL effects were lowly correlated between two populations (ρ = 0.2), and up to 0.16 when QTL effects were highly correlated (ρ = 0.8). When QTL genotypes were included for training and validation, the improvements were 0.16 and 0.22, respectively, for scenarios of the low and high correlation of QTL effects between two populations. When the number of QTL was large (n = 200), improvement was small with a maximum of 0.02 when QTL genotypes were not included for genomic prediction. Reduction in accuracy was observed for the simple pooling method when the number of QTL was small and correlation of QTL effects between the two populations was low. For the real data, the multi-task model achieved an increase of accuracy between 0 and 0.07 in the Ayrshire validation set when 28,206 SNPs were used, while the simple data pooling method resulted in a reduction of accuracy for all traits except for protein percentage. When 246,668 SNPs were used, the accuracy achieved from the multi-task model increased by 0 to 0.03, while using the pooling method resulted in a reduction of accuracy by 0.01 to 0.09. In the Holstein population, the three methods had similar performance.ConclusionsResults in this study suggest that the proposed multi-task Bayesian learning model for multi-population genomic prediction is effective and has the potential to improve the accuracy of genomic prediction.

Forest fires have important ecological, social and economic consequences causing loss of life and property. In order to prevent these consequences, it is very important to intervene in active fires in a timely manner and to determine the extent of burnt areas as soon as possible. In such studies, remote sensing methods provide great benefits in terms of speed and cost. In recent years, various methods have been developed to segment active fires and burnt areas with satellite images. Deep learning methods successfully perform segmentation processes in many areas such as disease detection in the field of health, crop type determination in the field of agriculture, land use and building detection in the field of urbanization. In this study, a method has been developed that automatically detects both active fires and burned areas that need to be re-enacted in terms of location and area size by using the same Sentinel 2 scene in a single time using deep learning methods. In particular, a new training and validation data set was created to train the U-Net+InceptionResNetV2 (CNN) model. By combining the powerful features of U-Net with InceptionResNet V2, a convolutional neural network trained over more than one million images on the ImageNet very base, we aim to examine its capabilities in burned area and active fire detection. The model applied on the test data has been shown to give successful results with an overall accuracy of 0.97 and an IoU (Intersection over union) value of 0.88 in the detection of burnt areas, and an overall accuracy of 0.99 and an IoU value of 0.82 in the detection of active fires. Finally, when the test images that were not used in the training dataset were evaluated with the trained model, it was revealed that the results were quite consistent in the detection of active fires and burnt areas and their geographical locations.

Assessment of VIIRS 375 m active fire detection product for direct burned area mapping

Agricultural straw burning is a significant source of greenhouse gas emissions, adversely affecting regional human health and air quality. Understanding the spatiotemporal patterns of agricultural fires is crucial for developing effective emissions reduction strategies in cropland to mitigate climate change. Although it is reported that cropland fires have been decreasing over the past two decades, the trends of global cropland fires on seasonal and diurnal scales remain poorly quantified, limiting a complete understanding of their spatiotemporal dynamics. This study analyzes global cropland fire activity from 2003 to 2020 at annual, seasonal, and diurnal scales, using multiple satellite-based burned area datasets, active fire products, and cropland classification datasets. The results show that from 2003 to 2020, global cropland burned area, active fire detections, and fire intensity all exhibited significant decreasing trends (p &lt; 0.05), with relative changes of -43.5%, -30.3%, and -3.5%, respectively. The most significant decreases in cropland burned area and active fire detections occurred in Africa, while the largest decline in fire intensity was observed in Asia. Moreover, cropland fire activity displayed notable seasonal and diurnal variations. On the seasonal scale, the largest declines in cropland burned area, active fire detections, and fire intensity were observed in December, August, and November, respectively. Notably, fire intensity showed a significant increasing trend (p &lt; 0.05) in April and September. On the diurnal scale, the decrease in cropland active fire detections was primarily driven by daytime activity; however, the rate of decline in fire intensity at night was about 1.5 times that during the day. These findings offer valuable insights into the comprehensive spatiotemporal patterns of global cropland fires, providing a foundation for more effective cropland management and carbon mitigation strategies.