Detection of Increased Erosion Damage Using Neural Networks on a Combination of Remote Sensing Imagery and Erosion-Hydrological Modeling

The project's goal is to create a software tool for detecting and predicting a higher form of (rill) erosion on agricultural land. The planned tool's innovative potential is the use of neural networks on the joint remote sensing and erosion-hydrological modelling data. Morphological parameters and erosion-hydrological causal event response thus enhance common inputs for the neural network-driven semantic segmentation.

By combining morphological parameters, event-based hydrological responses, and a calculated critical water layer thickness (h_crit) from physical SMODERP model - the threshold at which rill erosion begins - the tool enhances the precision of high-risk area delineation, supporting smart agriculture and climate adaptation.

The project utilizes a unique dataset of manually digitized erosion rills from over 20 years of aerial orthophotos, enabling comprehensive training of neural networks. Multi-resolution data, including satellite imagery, aerial orthophotos, and UAV images, are combined to identify and refine morphological properties critical for rill erosion detection. Several types of neural networks were tested, notably FCN, U-Net, SegNet, DeepLabv3+, to evaluate their effectiveness in handling diverse input data and optimizing predictive accuracy. Automated workflows for dataset expansion and retraining ensure adaptability to new data.

Validation of the model will be performed using the original dataset of manually digitized erosion rills as a benchmark for accuracy. By comparing the predicted rill locations with this dataset, the model’s performance can be rigorously evaluated and adjusted. Real-time erosion event mapping, supported by the Agricultural Land Erosion Monitoring system, will complement this process by incorporating contemporary data to further enhance model reliability. This innovative tool addresses gaps in existing methods by combining predictive capabilities with detailed spatial data, improving erosion detection accuracy for sustainable land management under changing climatic conditions.

The research is funded by the Technological Agency of the Czech Republic research project (TQ03000408)- Detection of Increased Erosion Damage Using Neural Networks on a Combination of Remote Sensing Imagery and Erosion-Hydrological Modeling and an internal student CTU grant (SGS23/155/OHK1/3T/11).