Machine Learning-Based Estimation of High-Resolution Snow Depth in Alaska Using Passive Microwave Remote Sensing Data

Snow Water Equivalent Monitoring—A Review of Large-Scale Remote Sensing Applications

“…Moreover, it becomes apparent that the commencement of coverage frequently coincides with the launch of specific sensors. In addition to the aforementioned SMMR in 1978 (n = 10), this mainly applies to the Special Sensor Microwave/Imager (SSM/I, n = 10) in 1987 [91,109,140,159,167,185,188,189,[194][195][196][197], the Advanced Microwave Scanning Radiometer for Earth Observation (AMSR-E, n = 27) [61,82,88,101,107,110,111,129,131,137,142,[150][151][152]155,156,158,162,[172][173][174][175]177,179,198,199], and the Advanced Microwave Scanning Radiometer 2 (AMSR-2, n = 4) [103,126,200,201]. The average coverage duration is just under 9.5 years with the longest study covering 41 years (1978-2019) [123], while some studies only examine a few months or one winter season…”

“…Within the AI method portion of the reviewed studies, the most frequently applied methods are SVM (n = 9), NN's (n = 7), and RF (n = 6). Regardless of the individual intentions and methods of the studies, most AI studies can be grouped together as they all use AI to uncover relationships between the Tb signal and other variables (such as, but not exclusively, hillslope, temperature, or precipitation) to improve SWE data [63,106,107,115,118,135,146,147,149,170,174,175,187,190,197,201,205,[229][230][231][232][233]. The remainder of the AI studies attempted to fuse existing datasets to fill data gaps and generate more robust data [121,125,133].…”

“…The individual tasks for which AI is used are diverse. For example, AI is employed to generate higher resolution [125,187,201,229,230], better depict SWE in mountain areas [187], or optimize existing methods [135,174,175].…”

“…One issue, thereby, is that several snow models do not depict all snow types correctly [246]. In more recent studies, researchers also used AI to address the snow grain size issue, especially by focusing on the sensitivity of the Tb signal using RF [231] or extremely randomized forests (ERFs) [201].…”

“…Within the group of studies that uses AI, RF is the predominant method (n = 6/7), be it applied on the the Ku-Ka combination [152], or ERF [201]. Wei published three studies comparing RF to other AI methods (SVR, XGBoost, and multiple linear regressions) [197,229,230].…”