Search

Convolutional neural networks (CNNs) are essential for image classification and detection, and their implementation in embedded systems is becoming increasingly attractive due to their compact size and low power consumption. Field-Programmable Gate Arrays (FPGAs) have emerged as a promising option, thanks to their low latency and high energy efficiency. Vitis AI and FINN are two development environments that automate the implementation of CNNs on FPGAs. Vitis AI uses a deep learning processing unit (DPU) and memory accelerators, while FINN is based on a streaming architecture and fine-tunes parallelization. Both environments implement parameter quantization techniques to reduce memory usage. This work extends previous comparisons by evaluating both environments by implementing four models with different numbers of layers on the Xilinx Kria KV260 FPGA platform. The complete process from training to evaluation on FPGA, including quantization and hardware implementation, is described in detail. The results show that FINN provides lower latency, higher throughput, and better energy efficiency than Vitis AI. However, Vitis AI stands out for its simplicity in model training and ease of implementation on FPGA. The main finding of this study is that as the complexity of the models increases (with more layers in the neural networks), the differences in terms of performance and energy efficiency between FINN and Vitis AI are significantly reduced.

In 2019 there were forest fires in Sumatra, Indonesia, which co-occurred with a strong positive IOD. The forest fire disaster caused the thick smoke containing dusts such as particulate matter with a size of 2.5μm (PM2.5). In this study, a model simulation was conducted to predict the distribution of PM2.5 using the WRF–Chem model when forest fires in 2019 occurred. The prediction was produced by adding the latest version of fire inventory from NCAR (FINN) input data, version 2.5. The prediction result was verified and compared with ground station data from the Ministry of Environment and Forestry (KLHK) station and NASA’s EOSDIS satellite imagery. There are three ground stations that were used for verification: the Jambi, Palembang and Pekanbaru stations. Of the three stations, the prediction results at the Palembang station are the best in correlation and RMSE value. Spatially, the distribution of PM2.5 from the model result is similar and can follow the pattern of smoke distribution from satellite imagery observations. Even though, generally the WRF–Chem model equipped with the latest FINN data still cannot produce an accurate prediction for the 2019 forest fires event yet in Sumatra region.

In the illustration accompanying ‘!CEÓL!’ (!Music!), Brian Ua Nualláin’s 1932 Irish-language story about radio and gramophone music, Ua Nualláin uses an established but outdated spelling of ‘ceol’ to suggest a polysemantic pun. This pun is significant given the story’s subtitle – ‘Eachtra an Fhir Ólta’ (The Tale of the Drunkard) – because it places ‘ól,’ the Irish word for drink, inside the word for music, ‘ceol.’ The resulting portmanteau in Ua Nualláin’s illustration is interfusional in Thomas King’s sense, because it blends the textual with the oral. The orthography Ua Nualláin uses for his visual representation of ‘!CEÓL!’ draws attention to how the word is spoken, and the stress on the second half of the word makes it sound drunken. We argue that the story’s implied title, ‘Drink-Music,’ thematises intertextual play in ways that anticipate Ua Nualláin’s Middle Irish Joycean pastiche ‘Pisa bec oc Parnabus’ (1935/1938). This note provides the first critical discussion of the ‘!CEÓL!’ illustration, observing the presence of Ua Nualláin’s initials in the bottom right-hand corner of the image.

Stubble-burning in northern India is an important source of atmospheric particulate matter (PM) and trace gases, which significantly impact local and regional climate, in addition to causing severe health risks. Scientific research on assessing the impact of these burnings on the air quality over Delhi is still relatively sparse. The present study analyzes the satellite-retrieved stubble-burning activities in the year 2021, using the MODIS active fire count data for Punjab and Haryana, and assesses the contribution of CO and PM2.5 from such biomass-burning activities to the pollution load in Delhi. The analysis suggests that the satellite-retrieved fire counts in Punjab and Haryana were the highest among the last five years (2016–2021). Further, we note that the stubble-burning fires in the year 2021 are delayed by ∼1 week compared to that in the year 2016. To quantify the contribution of the fires to the air pollution in Delhi, we use tagged tracers for CO and PM2.5 emissions from fire emissions in the regional air quality forecasting system. The modeling framework suggests a maximum daily mean contribution of the stubble-burning fires to the air pollution in Delhi in the months of October–November 2021 to be around 30–35%. We find that the contribution from stubble burning activities to the air quality in Delhi is maximum (minimum) during the turbulent hours of late morning to afternoon (calmer hours of evening to early morning). The quantification of this contribution is critical from the crop-residue and air-quality management perspective for policymakers in the source and the receptors regions, respectively.

Prescribed burning is a major source of a fine particular matter, especially in the southeastern United States, and quantifying emissions from burning operations accurately is an integral part of ascertaining air quality impacts. For instance, a critical factor in calculating fire emissions is identifying fire activity information (e.g., location, date/time, fire type, and area burned) and prior estimations of prescribed fire activity used for calculating emissions have either used burn permit records or satellite-based remote sensing products. While burn permit records kept by state agencies are a reliable source, they are not always available or readily accessible. Satellite-based remote sensing products are currently used to fill the data gaps, especially in regional studies; however, they cannot differentiate prescribed burns from the other types of fires. In this study, we developed novel algorithms to distinguish prescribed burns from wildfires and agricultural burns in a satellite-derived product, Fire INventory from NCAR (FINN). We matched and compared the burned areas from permit records and FINN at various spatial scales: individual fire level, 4 km grid level, and state level. The methods developed in this study are readily usable for differentiating burn type, matching and comparing the burned area between two datasets at various resolutions, and estimating prescribed burn emissions. The results showed that burned areas from permits and FINN have a weak correlation at the individual fire level, while the correlation is much higher for the 4 km grid and state levels. Since matching at the 4 km grid level showed a relatively higher correlation and chemical transport models typically use grid-based emissions, we used the linear regression relationship between FINN and permit burned areas at the grid level to adjust FINN burned areas. This adjustment resulted in a reduction in FINN-burned areas by 34%. The adjusted burned area was then used as input to the BlueSky Smoke Modeling Framework to provide long-term, three-dimensional prescribed burning emissions for the southeastern United States. In this study, we also compared emissions from different methods (FINN or BlueSky) and different data sources (adjusted FINN or permits) to evaluate uncertainties of our emission estimation. The comparison results showed the impacts of the burned area, method, and data source on prescribed burning emission estimations.

This study analyzed the characteristics of changes in O3 concentration in a plume induced by a wildfire in Andong, South Korea, from 24 to 26 April 2020, using the Community Multi-scale Air Quality (CMAQ) model. Fire INventory from National Center for Atmospheric Research (FINN) emissions data were used for the wildfire emissions. The increases in the concentrations of primary pollutants (CO, NOx, and volatile organic compounds (VOCs)) due to the wildfire peaked near the source at 09 LST and, as the plume was transported, the reduction in the supply of pollutants from wildfire, as well as chemical reactions, advection and diffusion, and deposition, caused the concentrations to continuously decrease. In contrast, O3 concentration showed a sustained increase during transport due to photochemical reactions caused by precursors (e.g., NOx, VOCs) emitted during the wildfire, peaking (1.40 ppb) at approximately 1 km at 13 LST over 60 km from the source. To analyze these results, a process analysis was conducted. Integrated process rate (IPR) analysis results showed that the production rate of O3 and loss rates of NOx and VOCs peaked at 09 LST due to the photochemical reactions of NOx and VOCs emitted due to wildfire. Then, as the plume was transported, the loss rates of NOx and VOCs that contributed to O3 production continued to decrease at 11 LST. The O3 production rate also decreased at 11 LST but increased at 13 LST due to increasing solar radiation. This indicates that the O3 concentration is complexly determined by O3 precursors and solar radiation. Additionally, IRR analysis revealed that NO and NO2 emitted during wildfire and solar radiation contributed to the production and loss processes of O3; the production reactions of O3 were predominant, and O3 was accumulated and transported in the plume, leading to the peak O3 concentration at 13 LST.

Object detection is an essential component of many systems used, for example, in advanced driver assistance systems (ADAS) or advanced video surveillance systems (AVSS). Currently, the highest detection accuracy is achieved by solutions using deep convolutional neural networks (DCNN). Unfortunately, these come at the cost of a high computational complexity; hence, the work on the widely understood acceleration of these algorithms is very important and timely. In this work, we compare three different DCNN hardware accelerator implementation methods: coarse-grained (a custom accelerator called LittleNet), fine-grained (FINN) and sequential (Vitis AI). We evaluate the approaches in terms of object detection accuracy, throughput and energy usage on the VOT and VTB datasets. We also present the limitations of each of the methods considered. We describe the whole process of DNNs implementation, including architecture design, training, quantisation and hardware implementation. We used two custom DNN architectures to obtain a higher accuracy, higher throughput and lower energy consumption. The first was implemented in SystemVerilog and the second with the FINN tool from AMD Xilinx. Next, both approaches were compared with the Vitis AI tool from AMD Xilinx. The final implementations were tested on the Avnet Ultra96-V2 development board with the Zynq UltraScale+ MPSoC ZCU3EG device. For two different DNNs architectures, we achieved a throughput of 196 fps for our custom accelerator and 111 fps for FINN. The same networks implemented with Vitis AI achieved 123.3 fps and 53.3 fps, respectively.

The Force Awakens enables, even fosters comparisons: is Rey a worthy hero? Is Finn believable as a once-Stormtrooper turned good resistance fighter? Is Kylo Ren anywhere close in power and ability to his role model, Darth Vader? Readers can glean my opinion from that evaluative sentence. Are the scenes parallel to earlier movies? This movie, the new characters, the new poster, the toys, the novelization (by Alan Dean Foster, the science fiction writer and film novelization specialist who served as ghost writer for George Lucas’ Star Wars: From the Adventures of Luke Skywalker, 1976, and Star Wars: The Force Awakens, 2016), and the countless viewing parties, the information sites such as Wookieepedia, and wonderful spoofs such as YouTube’s Bad Lip Reading enable me to say that this film is far beyond simple nostalgia and money-making scheme—though both of those are in play here. It is not the completeness and closure of The Force Awakens that compelled the editors to select that movie and those experiences for a special issue. Instead, it is the excitement that this jump-start has created. The structure at the center of the Star Wars universe has held, and the future looks bright for more new characters and also more wars. The battle for political and economic control on Jakku has something to do with our current Earth-bound problems about inequality and migration and war. But more locally, for Issue 12, questions about identity come to the fore: Who gets to create new Star Wars characters? Who gets to play those new characters? Who gets to play with storylines? Who gets to be the hero and villain? Who gets to voice opinions about the new shape of the First Order? These questions lead to “how” questions—especially how should new viewers and readers understand The Force Awakens. These questions motivate the four articles and one interview in this special issue.