Soil Parameters Prediction from Hyperspectral Images via Multitask Learning

Soil parameters are relevant and valuable for various applications such as agriculture production, scientific research, and policy making. Since acquiring such physical or chemical information could be cost-consuming by traditional methods, remote sensing and data analysis have become exciting research fields for soil parameter prediction tasks.   Many papers show that minerals and chemical materials correlate with the corresponding spectral reflectance. Based on this characteristic, there are many works using different data analysis methods such as machine learning and deep learning to predict soil properties, especially from multi- and hyper-spectral images. However, limited by the small data size, many models suffer from the overfitting problem and could not extrapolate to unseen data. Moreover, the currently existing methods only generate predictions with no consideration of the correlation among different target parameters. In this work, we propose and implement a deep learning based multitask method to predict multiple chemical properties simultaneously from hyperspectral images. To initialize the model, we use the pre-trained weights from ImageNet. To make better use of the correlation among different parameters, our model consists of shared layers and task-specific branches where each customized branch generates the prediction for one target property. Our method is implemented on the dataset from the Hyperview challenge organized by KP Labs and ESA. In this dataset, 1732 hyperspectral patches are available now and each patch has 4 soil parameters including K, P205, Mg, and pH. After comprehensive experiments, our method achieves the highest score of 0.87, which shows superior performance in this regression task.