Extending state-of-the-art object detectors from image to video is challenging. The accuracy of detection suffers from degenerated object appearances in videos, e.g., motion blur, video defocus, rare poses, etc. Existing work attempts to exploit temporal information on box level, but such methods are not trained end-to-end. We present flowguided feature aggregation, an accurate and end-to-end learning framework for video object detection. It leverages temporal coherence on feature level instead. It improves the per-frame features by aggregation of nearby features along the motion paths, and thus improves the video recognition accuracy. Our method significantly improves upon strong single-frame baselines in ImageNet VID [33], especially for more challenging fast moving objects. Our framework is principled, and on par with the best engineered systems winning the ImageNet VID challenges 2016, without additional bells-and-whistles. The proposed method, together with Deep Feature Flow [49], powered the winning entry of ImageNet VID challenges 2017 1 . The code is available at https://github.com/msracver/ Flow-Guided-Feature-Aggregation. * This work is done when Xizhou Zhu and Yujie Wang are interns at Microsoft Research Asia 1 http://image-net.org/challenges/LSVRC/2017/ results 1

Stomatal regulation presumably evolved to optimize CO for H O exchange in response to changing conditions. If the optimization criterion can be readily measured or calculated, then stomatal responses can be efficiently modelled without recourse to empirical models or underlying mechanism. Previous efforts have been challenged by the lack of a transparent index for the cost of losing water. Yet it is accepted that stomata control water loss to avoid excessive loss of hydraulic conductance from cavitation and soil drying. Proximity to hydraulic failure and desiccation can represent the cost of water loss. If at any given instant, the stomatal aperture adjusts to maximize the instantaneous difference between photosynthetic gain and hydraulic cost, then a model can predict the trajectory of stomatal responses to changes in environment across time. Results of this optimization model are consistent with the widely used Ball-Berry-Leuning empirical model (r > 0.99) across a wide range of vapour pressure deficits and ambient CO concentrations for wet soil. The advantage of the optimization approach is the absence of empirical coefficients, applicability to dry as well as wet soil and prediction of plant hydraulic status along with gas exchange.

Animals living in the extremely cold environment, such as polar bears, have shown amazing capability to keep warm, benefiting from their hollow hairs. Mimicking such a strategy in synthetic fibers would stimulate smart textiles for efficient personal thermal management, which plays an important role in preventing heat loss and improving efficiency in house warming energy consumption. Here, a "freeze-spinning" technique is used to realize continuous and large-scale fabrication of fibers with aligned porous structure, mimicking polar bear hairs, which is difficult to achieve by other methods. A textile woven with such biomimetic fibers shows an excellent thermal insulation property as well as good breathability and wearability. In addition to passively insulating heat loss, the textile can also function as a wearable heater, when doped with electroheating materials such as carbon nanotubes, to induce fast thermal response and uniform electroheating while maintaining its soft and porous nature for comfortable wearing.

No abstract

Abstract. This paper describes the latest version of the algorithm MAIAC used for processing the MODIS Collection 6 data record. Since initial publication in 2011–2012, MAIAC has changed considerably to adapt to global processing and improve cloud/snow detection, aerosol retrievals and atmospheric correction of MODIS data. The main changes include (1) transition from a 25 to 1 km scale for retrieval of the spectral regression coefficient (SRC) which helped to remove occasional blockiness at 25 km scale in the aerosol optical depth (AOD) and in the surface reflectance, (2) continuous improvements of cloud detection, (3) introduction of smoke and dust tests to discriminate absorbing fine- and coarse-mode aerosols, (4) adding over-water processing, (5) general optimization of the LUT-based radiative transfer for the global processing, and others. MAIAC provides an interdisciplinary suite of atmospheric and land products, including cloud mask (CM), column water vapor (CWV), AOD at 0.47 and 0.55 µm, aerosol type (background, smoke or dust) and fine-mode fraction over water; spectral bidirectional reflectance factors (BRF), parameters of Ross-thick Li-sparse (RTLS) bidirectional reflectance distribution function (BRDF) model and instantaneous albedo. For snow-covered surfaces, we provide subpixel snow fraction and snow grain size. All products come in standard HDF4 format at 1 km resolution, except for BRF, which is also provided at 500 m resolution on a sinusoidal grid adopted by the MODIS Land team. All products are provided on per-observation basis in daily files except for the BRDF/Albedo product, which is reported every 8 days. Because MAIAC uses a time series approach, BRDF/Albedo is naturally gap-filled over land where missing values are filled-in with results from the previous retrieval. While the BRDF model is reported for MODIS Land bands 1–7 and ocean band 8, BRF is reported for both land and ocean bands 1–12. This paper focuses on MAIAC cloud detection, aerosol retrievals and atmospheric correction and describes MCD19 data products and quality assurance (QA) flags.

No abstract

A number of models have been developed to estimate PM2.5 exposure, including satellite-based aerosol optical depth (AOD) models, land-use regression or chemical transport model simulation, all with both strengths and weaknesses. Variables like normalized difference vegetation index (NDVI), surface reflectance, absorbing aerosol index and meteoroidal fields, are also informative about PM2.5 concentrations. Our objective is to establish a hybrid model which incorporates multiple approaches and input variables to improve model performance. To account for complex atmospheric mechanisms, we used a neural network for its capacity to model nonlinearity and interactions. We used convolutional layers, which aggregate neighboring information, into a neural network to account for spatial and temporal autocorrelation. We trained the neural network for the continental United States from 2000 to 2012 and tested it with left out monitors. Ten-fold cross-validation revealed a good model performance with total R2 of 0.84 on the left out monitors. Regional R2 could be even higher for the Eastern and Central United States. Model performance was still good at low PM2.5 concentrations. Then, we used the trained neural network to make daily prediction of PM2.5 at 1 km×1 km grid cells. This model allows epidemiologists to access PM2.5 exposure in both the short-term and the long-term.

We show that the vegetation canopy of the Amazon rainforest is highly sensitive to changes in precipitation patterns and that reduction in rainfall since 2000 has diminished vegetation greenness across large parts of Amazonia. Large-scale directional declines in vegetation greenness may indicate decreases in carbon uptake and substantial changes in the energy balance of the Amazon. We use improved estimates of surface reflectance from satellite data to show a close link between reductions in annual precipitation, El Niño southern oscillation events, and photosynthetic activity across tropical and subtropical Amazonia. We report that, since the year 2000, precipitation has declined across 69% of the tropical evergreen forest (5.4 million km 2 ) and across 80% of the subtropical grasslands (3.3 million km 2 ). These reductions, which coincided with a decline in terrestrial water storage, account for about 55% of a satellite-observed widespread decline in the normalized difference vegetation index (NDVI). During El Niño events, NDVI was reduced about 16.6% across an area of up to 1.6 million km 2 compared with average conditions. Several global circulation models suggest that a rise in equatorial sea surface temperature and related displacement of the intertropical convergence zone could lead to considerable drying of tropical forests in the 21st century. Our results provide evidence that persistent drying could degrade Amazonian forest canopies, which would have cascading effects on global carbon and climate dynamics.Amazon | climate change | precipitation | NDVI | MODIS | MAIAC