Temporal and Spatial Analysis of Critical Field Points Using High-Resolution Soil Water Content Estimation Employing Remote Sensing and Deep-Learning

This study presents a novel deep-learning approach for estimating Soil Water Content (SWC) with high spatial resolution across multiple soil depths. Additionally, the study identifies critical field points based on their drying-out times analyzed by SWC estimations over extended periods. Understanding potential critical points regarding SWC allows operators of heavy agricultural equipment to gain insight into the field's traits and prevent excessive soil compaction. Additionally, this information can support more strategic and efficient harvesting plans by accounting for the impact of varying drying patterns on crop growth and soil strength to not only minimize soil degradation but also maximize yield production, offering a more productive and sustainable crop production.

In this regard, our proposed method offers a practical approach to integrating diverse data types, including:

• Spatial data: remote sensing data (Synthetic Aperture Radar (SAR) and vegetation index), land elevation, and soil profiles at various depths (soil content and bulk density).
• Temporal data: historical weather information (precipitation, temperature, wind, and global radiation).
• Contextual data: date, groundwater level, and crop type.

Previous machine learning and numerical models primarily used temporal and contextual data alongside point-based parameter values as inputs. In contrast, we incorporated spatial information instead of point values, allowing the model to capture better the surrounding influences—such as elevation, water flow, and vegetation shadows—on SWC.

To be able to estimate the SWC using the comprehensive analysis of spatial, temporal, and relevant contextual factors, these inputs are processed by a novel multi-model deep learning framework comprising:

• U-Net to capture spatial features and the impacts of 2D image data.
• Temporal Convolutional Network (TCN) to extract temporal dependencies from weather data.
• Feed-Forward Network (FFN) to model the influence of contextual inputs.

Our model is trained and validated using ground truth data from site measurements in the HOBE dataset. These measurements are conducted at 30 locations within the Skjern River Catchment in Western Denmark, with each data sample containing SWC at different depths: surface, 20cm, and 50cm. By utilizing data collected between 2014 and 2018 from point 1.09 in the HOBE dataset, we demonstrated that the proposed model achieved a mean absolute error (MAE) of 0.0207. For comparison, a numerical model (Daisy) and a machine learning approach that did not account for spatial context produced higher MAEs of 0.0382 and 0.0269, respectively.

Subsequently, the developed model is employed to estimate SWC over extended periods and identify critical points within fields. To achieve this, we collaborated with several farmers who manually classified their field maps into regular, late-drying, and critical parts. The distinction between the latter two categories is crucial, as our observations revealed that "not every wet point is a critical point." The collected temporal SWC data is analyzed with land elevation to differentiate between these two classes. This aspect of the study remains under investigation, and further research is being conducted to refine the classification process and validate its effectiveness.