SLVVA: Scalable Land Viability via Vision-Language Architecture

Digital soil mapping is a process of creating maps of soil properties and their spatial distribution. It plays a vital role in monitoring soil health and promoting sustainable and efficient land use. In the past, environmental data was used to guide the creation of soil property maps. However, the failure to consider the accessibility of locations has led to a bias in the mapping process. In our research, we utilize satellite imagery to evaluate location accessibility, leading to more balanced soil property mapping. We formulate land viability detection and introduce a scalable two-step framework for its detection. Initially, we classify land viability, followed by its segmentation. We leverage Convolutional Neural Networks (CNNs) for classification and a resilient and generalizable vision-language architecture for segmentation. Our most notable results stem from fine-tuning a pre-existing VGGNet for classification and employing a CLIP-based Segmentation method (CLIPSeg) for segmentation. We demonstrate the effectiveness of our approach through extensive experimentation on EuroSAT and OpenEarthMap datasets. Our work is the first to address the challenge of biased sampling in digital soil mapping by incorporating satellite images to assess the accessibility of locations, ensuring a more representative soil property mapping.