DEEP AND SHALLOW LEARNING IN DEEP AND SHALLOW SNOW: SNOWNET

THE ABILITY TO ACCURATELY QUANTIFY SNOWPACK VARIABILITY HAS MAJOR IMPACTS ON ESTIMATIONS OF SNOW DEPTH SNOW WATER EQUIVALENT (SWE) AND ULTIMATELY NET WATER STORAGE. COMPLEX PHYSICS BASED MODELS HAVE ATTEMPTED TO CAPTURE THE PRIMARY PHYSICAL MECHANISMS THAT AFFECT SNOWPACK PROPERTIES AND HAVE BEEN CAPABLE OF ESTIMATING SNOW PROPERTIES ACCURATELY IF DRIVEN BY ACCURATE METEOROLOGICAL FORCINGS. HOWEVER THEY TYPICALLY FAIL TO MODEL ABRUPT CHANGES IN SNOWPACK ESPECIALLY IN MOUNTAINOUS TERRAIN AND DO NOT MODEL SUBKILOMETER SNOW VARIABILITY WELL. THIS IS IN PART DUE TO THE LACK OF IN-SITU MEASUREMENTS THAT CONTAIN THE DETAILED FORCING DATA REQUIRED BY THE COMPLEX MODELS COMBINED WITH OUR POOR UNDERSTANDING OF THE RELATIONSHIP BETWEEN MEASURED AND UNMEASURED LOCATIONS. THE WORK PROPOSED HERE WILL LEVERAGE NEWLY AVAILABLE DATASETS FROM: THE NASA SNOWEX INSITU MEASUREMENTS LEGACY DATASETS FROM: NASA CLPX I BOISE DRY CREEK AND REYNOLDS CREEK EXPERIMENTAL WATERSHEDS; AND PHYSICAL MODEL OUTPUTS FROM SOLAR RADIATION AND WIND VECTOR FIELDS. OUR APPROACH TO BETTER UNDERSTAND SNOW DISTRIBUTION IS THROUGH THE EYES OF A DATA AND MACHINE-LEARNING SCIENTIST. THE RISE OF COMPUTATIONAL POWER HAS OPENED THE DOOR TO EFFICIENT EXPLORATION OF LARGE DATASETS. MOREOVER NEWLY DEVELOPED ARTIFICIAL INTELLIGENCE (AI) ALGORITHMS HAVE BEGUN TO SURPASS HUMAN PERFORMANCE SPECIFICALLY IN IMAGE AND SEQUENTIAL DATA. WE PROPOSE TO USE AI TO EXPLORE THE NONLINEAR SYSTEMS WITHIN THE CRYOSPHERE IN HOPES TO BETTER UNDERSTAND HOW THE SEASONAL SNOWPACK IS CHANGING SPATIALLY AND TEMPORALLY. THIS PROJECT WILL USE REAL TIME APPLICATIONS OF AI WHICH WILL REQUIRE EFFICIENTLY STORED ARCHIVES OF DATA. WE PROPOSE TO UTILIZE GRAPHS TO STORE DATA THOSE DATA FOR FAST QUERIES; GRAPH DATABASES ARE IDEAL FOR SUCH REAL-TIME TASKS. INCLUSIVELY GRAPHS HAVE UNIQUE PROPERTIES THAT ALLOW DELINEATION OF OPTIMAL IN-SITU SITES AND TRAINING SETS FOR MACHINE-DRIVEN ALGORITHMS. THE OBJECTIVE OF THIS WORK IS TO: 1) DEFINE A CLEAR METHOD TO MANAGE LARGE CRYOSPHERE DATASETS IN GRAPH DATABASE AS WELL AS TECHNIQUES ON EFFICIENT QUERIES 2) IMPLEMENT OF VARIOUS MACHINE LEARNING ALGORITHMS TO MAXIMIZE INFORMATION RETRIEVAL FROM A VARIETY OF REMOTELY SENSED AND INSITU MEASUREMENTS AND 3) IDENTIFY REGIONAL AND LOCAL IN-SITU MEASUREMENTS OF SNOWPACK CONDITIONS. THE RESULTS OF THIS WORK WILL PROVIDE BETTER WATER SUPPLY ESTIMATES USING STATE-OF-THE-ART TECHNOLOGY AND REMOTE SENSING PLATFORMS AND WILL ALLOW FOR DETECTION OF KEY VALIDATION POINTS FOR FUTURE MONITORING SITES TO IMPROVE CURRENT AND FUTURE SPRING/SUMMER RUN-OFF MODELS. THIS PROJECT WILL PRODUCE A WELL-DOCUMENTED IMPLEMENTATION OF NASA PROVIDED DATASETS VIA GRAPHICAL MODELS DEEP CONVOLUTIONAL AND RECURRENT NEURAL NETWORKS THROUGH HIGHLY COMPUTATIONAL EFFICIENT OPEN SOURCE PLATFORMS SUCH AS NEO4J TENSORFLOW THEANO AND KERAS TO PREDICT SNOW DEPTH--A HIGHLY VARIABLE UNKNOWN STATE NEEDED FOR ESTIMATIONS OF FROZEN WATER STORAGE IN MOUNTAINOUS TERRAIN.