Mapping Soil Organic Matter Under Field Conditions

Soil Organic Matter (SOM) is a key component for sustainable agriculture planning and soil management. Nutrient analysis, spectroscopy and digital soil imaging are commonly used to estimate SOM in a controlled lab setting. These methods are accurate, but the controlled lab setting is not scalable. For scalability, high-resolution satellite imagery is widely employed. However, special conditions of the Canadian Prairies, like harsh weather and crop residue cover, pose significant challenges in getting the spectral signatures of bare soil. To overcome these challenges, this paper presents a novel methodology that explores the prospects of using high-resolution ground images acquired under Uncontrolled Field Conditions (UFC) for SOM estimation. The developed methodology first extracts bare soil from images using deep learning methods. As the image samples are acquired under UFC, variable ambient illumination influences soil colour. To counter this, in the second step, we propose unsupervised colour constancy to mitigate the effects of variable ambient lighting conditions. In the third step, colour space and texture features are extracted to estimate SOM. We compare our proposed method with the state-of-the-art (SOTA) SOM estimation methods. We also performed an ablation study to compare the results of the SOTA with and without the addition of the colour constancy block. With the developed methodology, our bare soil segmentation model achieves a mean intersection over union value of 0.8134. Similarly, with the colour constancy methods applied on bare soil segmented images, our proposed method improves the
$R^{2}$
score by more than 30% with respect to the SOTA.