Soil Moisture Content Retrieval from Remote Sensing Data by Artificial Neural Network Based on Sample Optimization

Qixin Liu, Xingfa Gu, Xinran Chen, Faisal Mumtaz, Yan Liu, Chunmei Wang, Tao Yu, Yin Zhang, Dakang Wang, Yulin Zhan2022 · DOI

TL;DR

This paper introduces a novel "Sparse Sample Exploitation" (SSE) method to overcome data scarcity in remote sensing for soil moisture retrieval, significantly increasing the number of usable samples. By combining this with an expanded set of input parameters for an Artificial Neural Network (ANN), they achieve improved soil moisture retrieval accuracy, demonstrating the importance of both data quantity and quality.

What Problem Does This Paper Address?

Accurately estimating soil moisture content (SMC) is crucial for various geoscience applications, from climate change prediction to hydrology and agriculture. While remote sensing (RS) offers a powerful way to monitor SMC over large areas, traditional methods face challenges. Optical RS data is often hampered by cloud cover, and in-situ measurements can be sparse and prone to instrument malfunctions, leading to a shortage of complete, high-quality data for training machine learning models like Artificial Neural Networks (ANNs). This data deficiency, both in quantity and quality of input features, limits the accuracy of SMC retrieval.

Methods

The core innovation in this paper is the Sparse Sample Exploitation (SSE) method. This technique addresses the "data scarcity" problem by intelligently incorporating samples that would typically be discarded. Instead of requiring complete data (optical, microwave, and in-situ) for every monitoring site on a given day, SSE allows for the inclusion of samples where some data might be missing (e.g., due to cloud cover in optical images or absent in-situ measurements), as long as other essential data points are available. This effectively expands the training dataset.

Beyond data quantity, the paper also focuses on sample quality by optimizing the input parameters for the ANN. While common parameters like radar backscattering coefficients (VH/VV) and NDVI were used, the researchers explored the impact of additional parameters: data acquisition time (month), land surface temperature (LST), elevation, slope, and land cover type. They systematically tested various combinations of these parameters to understand their influence on SMC retrieval accuracy.

The retrieval itself is performed using an Artificial Neural Network (ANN), specifically a feed-forward perceptron model with a 3-layer structure (input, hidden, output). The ANN is trained using the expanded sample pool and optimized input parameters.

Datasets

  • Ground Truth Data: Soil moisture content (SMC) data from the International Soil Moisture Network (ISMN), specifically from the WEGENERNET (9 sites) and GROW (13 sites) networks in Austria, collected at a depth of 0-5 cm.
  • Remote Sensing Data:
    • Optical: Landsat-8 Operational Land Imager (OLI) and Thermal Infrared Sensor (TIRS) images.
  • Microwave: Sentinel-1 Synthetic Aperture Radar (SAR) images (IW mode, C-band, VH and VV polarization).
  • Auxiliary Data:
    • Digital Elevation Model (DEM) from Shuttle Radar Topography Mission (SRTM).
  • Land Cover product: "Global Land Cover with Fine Classification System at 30 m" (GLC_FCS30).

The paper does not introduce new benchmark datasets but rather utilizes existing ground truth and remote sensing data, applying a novel method to optimize their use.

Key Insights and Ideas

  • The power of "imperfect" data: The SSE method demonstrates that even "incomplete" data points, when intelligently handled, can significantly contribute to improving model performance, challenging the strict data completeness requirements of traditional approaches.
  • Beyond the basics: Simply using standard remote sensing inputs for SMC retrieval is insufficient. Incorporating ancillary data like elevation and slope, which capture local topographical influences on water movement, can substantially boost accuracy.
  • Elevation is king (for this study): Among the tested parameters, elevation emerged as the most influential factor for improving SMC retrieval accuracy, highlighting the critical role of topography in hydrological processes.
  • Land cover nuance: While land cover is generally important, its effectiveness as an input parameter can be limited by the resolution and accuracy of the land cover map itself, as seen in this study where a coarse categorization yielded minimal improvement.

Key Results

  • The SSE method successfully increased the number of available samples from 264 (traditional method) to 635.
  • With the SSE method and all nine input parameters, the testing dataset's correlation coefficient (r) increased from 0.635 to 0.736, and the Root Mean Square Error (RMSE) decreased from 0.090 m³m⁻³ to 0.068 m³m⁻³.
  • Incorporating elevation as an input parameter led to the most significant improvement, with the r-value increasing from 0.588 to 0.689 and RMSE decreasing from 0.089 m³m⁻³ to 0.070 m³m⁻³.
  • The study achieved an overall correlation coefficient of 0.84972 and an RMSE of 0.052 m³m⁻³ for the testing dataset when using all nine parameters and the SSE method.

Limitations and Future Work

  • Inconsistency between ground truth and RS data: The study acknowledges that the geographical distance of the study area from the researchers and the lack of field surveys for topography and land cover might introduce inconsistencies, leading to potential biases in SMC retrieval.
  • Land cover categorization: The limited improvement from the land cover type parameter suggests that a more refined land cover map is needed for better results.
  • Future Work: The authors plan to explore sample exploitation in the spatial dimension, using data from similar geographical sites with sufficient samples for areas lacking them. They also aim to incorporate data from newer Chinese satellites (GF-3 and GF-1) for SMC retrieval.

Summary updated

BibTeX
@misc{liu2022soil,
  author = {Qixin Liu and Xingfa Gu and Xinran Chen and Faisal Mumtaz and Yan Liu and Chunmei Wang and Tao Yu and Yin Zhang and Dakang Wang and Yulin Zhan},
  title = {{Soil Moisture Content Retrieval from Remote Sensing Data by Artificial Neural Network Based on Sample Optimization}},
  year = {2022},
  doi = {10.3390/s22041611},
}
Figure