In this paper, to break the limit of the traditional linear models for synthetic aperture radar (SAR) image despeckling, we propose a novel deep learning approach by learning a non-linear end-to-end mapping between the noisy and clean SAR images with a dilated residual network (SAR-DRN). SAR-DRN is based on dilated convolutions, which can both enlarge the receptive field and maintain the filter size and layer depth with a lightweight structure. In addition, skip connections and a residual learning strategy are added to the despeckling model to maintain the image details and reduce the vanishing gradient problem. Compared with the traditional despeckling methods, the proposed method shows a superior performance over the state-of-the-art methods in both quantitative and visual assessments, especially for strong speckle noise.

Accurate mapping of farmland soil organic carbon (SOC) provides valuable information for evaluating soil quality and guiding agricultural management. The integration of natural factors, agricultural activities, and landscape patterns may well fit the high spatial variation of SOC in low-relief farmlands. However, commonly used prediction methods are global models, ignoring the stratified heterogeneous relationship between SOC and environmental variables and failing to reveal the determinants of SOC in different subregions. Using 242 topsoil samples collected from Jianghan Plain, China, this study explored the stratified heterogeneous relationship between SOC and natural factors, agricultural activities, and landscape metrics, determined the dominant factors of SOC in each stratum, and predicted the spatial distribution of SOC using the Cubist model. Ordinary kriging, stepwise linear regression (SLR), and random forest (RF) were used as references. SLR and RF results showed that land use types, multiple cropping index, straw return, and percentage of water bodies are global dominant factors of SOC. Cubist results exhibited that the dominant factors of SOC vary in different cropping systems. Compared with the SOC of paddy fields, the SOC of irrigated land was more affected by irrigation-related factors. The effect of straw return on SOC was diverse under different cropping intensities. The Cubist model outperformed the other models in explaining SOC variation and SOC mapping (fitting R² = 0.370 and predicted R² = 0.474). These results highlight the importance of exploring the stratified heterogeneous relationship between SOC and covariates, and this knowledge provides a scientific basis for farmland zoning management. The Cubist model, integrating natural factors, agricultural activities, and landscape metrics, is effective in explaining SOC variation and mapping SOC in low-relief farmlands.

In this paper, to break the limit of the traditional linear models for synthetic aperture radar (SAR) image despeckling, we propose a novel deep learning approach by learning a non-linear end-to-end mapping between the noisy and clean SAR images with a dilated residual network (SAR-DRN). SAR-DRN is based on dilated convolutions, which can both enlarge the receptive field and maintain the filter size and layer depth with a lightweight structure. In addition, skip connections and residual learning strategy are added to the despeckling model to maintain the image details and reduce the vanishing gradient problem. Compared with the traditional despeckling methods, the proposed method shows superior performance over the state-of-the-art methods on both quantitative and visual assessments, especially for strong speckle noise.