Novel Dry Soil and Vegetation Indices to Predict Soil Contents from Landsat 8 Satellite Data.

Soil content prediction from satellite image data is an emerging research field as it can handle vast land areas non-invasively. However, this approach faces challenges due to the moisture and soil/vegetation patterns, as the hyperspectral reflections are varied due to those phenomena, although the soil content may be the same. To handle this, soil and vegetation indices are already proposed; however, there are no explicit indices to handle dry soil and dry vegetation. The existing soil index cannot separate dry soil from moist soil; on the other hand, the existing vegetation index cannot separate dead vegetation from bare soil. This paper, to the best of our knowledge, is the first to introduce dry soil index (DSI) and dry vegetation index (DVI) by considering the hyperspectral reflection variations and different moisture levels. Unlike the existing indices, the proposed indices use non-visual range bands in the formulations due to the greater reflection variations. This study uses the publicly available Land Use and Coverage Area Body Survey (LUCAS) 2018 dataset and corresponding Landsat 8 satellite images. We have used a number of learning-based models to analyze the performance, and we use DVI and DSI as features for predicting three soil contents (i.e., organic carbon, calcium carbonate, and nitrogen). Experimental results show that DVI and DSI are essential in filtering bare soil to improve soil content prediction performance.