Plastic waste is a significant environmental pollutant that is difficult to monitor. We created a system of neural networks to analyze spectral, spatial, and temporal components of Sentinel-2 satellite data to identify terrestrial aggregations of waste. The system works at wide geographic scale, finding waste sites in twelve countries across Southeast Asia. We evaluated performance in Indonesia and detected 374 waste aggregations, more than double the number of sites found in public databases. The same system deployed in Southeast Asia identifies 996 subsequently confirmed waste sites. For each detected site, we algorithmically monitor waste site footprints through time and cross-reference other datasets to generate physical and social metadata. 19% of detected waste sites are located within 200 m of a waterway. Numerous sites sit directly on riverbanks, with high risk of ocean leakage.

Land use/land cover (LULC) maps are foundational geospatial data products needed by analysts and decision makers across governments, civil society, industry, and finance to monitor global environmental change and measure risk to sustainable livelihoods and development. There is a strong need for high-level, automated geospatial analysis products that turn these pixels into actionable insights for non-geospatial experts. The Sentinel 2 satellites, first launched in mid-2015, are excellent candidates for LULC mapping due to their high spatial, spectral, and temporal resolution. Advances in deep learning and scalable cloud-based compute now provide the analysis capability required to unlock the value in global satellite imagery observations. Based on a novel, very large dataset of over 5 billion human-labeled Sentinel-2 pixels, we developed and deployed a deep learning segmentation model on Sentinel-2 data to create a global LULC map at 10m resolution that achieves state-of-the-art accuracy and enables automated LULC mapping from time series observations.

We outline the core components of a modulation recognition system that uses hierarchical deep neural networks to identify data type, modulation class and modulation order. Our system utilizes a flexible front-end detector that performs energy detection, channelization and multi-band reconstruction on wideband data to provide raw narrowband signal snapshots. We automatically extract features from these snapshots using convolutional neural network layers, which produce decision class estimates. Initial experimentation on a small synthetic radio frequency dataset indicates the viability of deep neural networks applied to the communications domain. We plan to demonstrate this system at the Battle of the Mod Recs Workshop at IEEE DySpan 2017.