Preprint Article Version 1 This version is not peer-reviewed

Localized Learning of Downscaled Soil Moisture

Version 1 : Received: 4 February 2019 / Approved: 5 February 2019 / Online: 5 February 2019 (08:01:58 CET)

How to cite: Lewis, M.; Fisher, A.; Smith, C.; Qu, J.; Houser, P. Localized Learning of Downscaled Soil Moisture. Preprints 2019, 2019020046 (doi: 10.20944/preprints201902.0046.v1). Lewis, M.; Fisher, A.; Smith, C.; Qu, J.; Houser, P. Localized Learning of Downscaled Soil Moisture. Preprints 2019, 2019020046 (doi: 10.20944/preprints201902.0046.v1).

Abstract

If given the correct remotely sensed information, machine learning can accurately describe soil moisture conditions in a heterogeneous region at the large scale based on soil moisture readings at the small scale through rule transference across scale. This paper reviews an approach to increase soil moisture resolution over a sample region over Australia using the Soil Moisture Active Passive (SMAP) sensor and Landsat 8 only and a validation experiment using Sentinal-2 and the Advanced Microwave Scanning Radiometer (AMSR-E) over Nevada. This approach uses an inductive localized approach, replacing the need to obtain a deterministic model in favor of a learning model. This model is adaptable to heterogeneous conditions within a single scene unlike traditional polynomial fitting models and has fixed variables unlike most learning models. For the purposes of this analysis, the SMAP 36 km soil moisture product is considered fully valid and accurate. Landsat bands coinciding in collection date with a SMAP capture are down sampled to match the resolution of the SMAP product. A series of indices describing the Soil-Vegetation-Atmosphere Triangle (SVAT) relationship are then produced, including two novel variables, using the down sampled Landsat bands. These indices are then related to the local coincident SMAP values to identify a series of rules or trees to identify the local rules defining the relationship between soil moisture and the indices. The defined rules are then applied to the Landsat image in the native Landsat resolution to determine local soil moisture. Ground truth comparison is done via a series of grids using point soil moisture samples and air-borne L-band Multibeam Radiometer (PLMR) observations done under the SMAPEx-5 campaign. This paper uses a random forest due to its highly accurate learning against local ground truth data yet easily understandable rules. The predictive power of the inferred learning soil moisture algorithm did well with a mean absolute error of 0.054 over an airborne L-band retrieved surface over the same region.

Subject Areas

Soil Moisture; Remote Sensing; Landsat; SMAP; Random Forest; Machine Learning; Downscaling; Microwave

Comments (0)

We encourage comments and feedback from a broad range of readers. See criteria for comments and our diversity statement.

Leave a public comment
Send a private comment to the author(s)
Views 0
Downloads 0
Comments 0
Metrics 0


×
Alerts
Notify me about updates to this article or when a peer-reviewed version is published.