DEEP LEARNING OF DRONE IMAGES FOR SOIL EROSION DETECTION USING GEOLOGICAL AND TOPOGRAPHIC LANDCOVER FEATURES

Soil erosion is a major threat to arable land and wildlife habitat, particularly in southern Africa, where a semiarid climate is paired with intensive stock grazing. The South African term “donga” is used for erosional features in semiarid areas that manifest as dry gullies formed by the action of ephemeral running water. Semi-arid countries such as South Africa are at the greatest risk of heavy erosion caused by geological formations such as dongas. These formations need to be monitored and managed but determining the presence of dongas is an expensive and resource intensive task. Drone surveys offer an economical alternative, but an approach is required to automatically classify these features. Extending on an earlier version of a land classifier, a model has been developed that is computationally light compared to very deep learning models and was developed with mobile hardware in mind. A patch level approach was used to maximize the data available during training. An improved model has been developed with superior performance, that generalizes well and is less error prone. The hyperspectral dataset consists of features including spectral values, height, aspect and slope. Explainable AI via was performed on the proposed model to try and shedding light on the reasoning behind the models’ outputs. This was achieved via an in-depth error analysis and saliency maps to visualize the attention of the proposed model per patch. Using the hyperspectral dataset, the proposed model reported a F1-Score of 0,948 and a Cohens Kappa of 0,935. This model outcompeted ResNet50, VGG16, EfficientNet-B0 and EfficientNet-B4 by several points. The performance of the proposed model is state-of-the-art and this resulting land classifier model will make an excellent tool for determining high erosion areas within a drone-based image map.