Background: The compaction of subsoils in agriculture is a threat to soil functioning. Measures aimed at the prevention, amelioration, and/or impact alleviation of compacted subsoils have been studied for more than a century, but less in smallholder agriculture. Methods: A meta-analysis was conducted to quantitatively examine the effects of the prevention, amelioration, and impact alleviation measures in mechanized and small-holder agriculture countries, using studies published during 2000~2019/2020. Results: Mean effect sizes of crop yields were large for controlled traffic (+34%) and irrigation (+51%), modest for subsoiling, deep ploughing, and residue return (+10%), and negative for no-tillage (−6%). Mean effect sizes of soil bulk density were small (<10%), suggesting bulk density is not a sensitive ‘state’ indicator. Mean effect sizes of penetration resistance were relatively large, with large variations. Controlled traffic had a larger effect in small-holder farming than mechanized agriculture. Conclusion: We found no fundamental differences between mechanized and smallholder agriculture in the mean effect sizes of the prevention, amelioration, and impact alleviation measures. Measures that prevent soil compaction are commonly preferred, but amelioration and alleviation are often equally needed and effective, depending on site-specific conditions. A toolbox of soil compaction prevention, amelioration, and alleviation measures is needed, for both mechanized and smallholder agriculture.

Current tillage practices in the important winter wheat–summer maize double cropping system of the North China Plain are under debate because of negative effects on soil quality and crop yield. Therefore, a long-term experiment was conducted from 2001 to 2018 to determine the effects of soil conservation practices on crop yield and soil quality. The treatments were imposed following maize harvest and prior wheat seeding, and were defined as follows: (1) moldboard ploughing (0–20 cm) following maize straw removal (CK); (2) moldboard ploughing (0–20 cm) following maize straw return (CT); (3) rotary tillage following maize straw return (RT); and (4) no tillage with maize straw covering the soil surface (NT). Wheat straw was chopped and spread on the soil in all treatments and maize seeded without prior tillage. Wheat yields were higher in CT than RT and NT treatments (p < 0.05); NT had 18% lower wheat yields than CT. No significant differences were found between treatments in summer maize yields. The soil organic carbon (SOC) content in the surface layer (0–5 cm) was higher in NT and RT compared to CT and CK. However, SOC content in the 10–20 cm and 20–30 cm layers was lower in NT and RT compared to CT and CK. Similarly, available phosphorus in the surface soil was higher in NT and RT than in CT and CK. but the opposite was true for the lower soil layers. SOC stocks (0–30 cm) increased in all treatments, and were initially faster in NT and RT than in CT and CK. However, SOC stocks were higher in CT than in other treatments at the end of the experiment. This finding indicates that no tillage and reduced tillage decreased both wheat yields and soil C sequestration over time; it also indicates that CT was the most robust in terms of crop yields and soil C sequestration.

Extracting land cover information from satellite imagery is of great importance for the task of automated monitoring in various remote sensing applications. Deep convolutional neural networks make this task more feasible, but they are limited by the small dataset of annotated images. In this paper, we present a fully convolutional networks architecture, FPN-VGG, that combines Feature Pyramid Networks and VGG. In order to accomplish the task of land cover classification, we create a land cover dataset of pixel-wise annotated images, and employ a transfer learning step and the variant dice loss function to promote the performance of FPN-VGG. The results indicate that FPN-VGG shows more competence for land cover classification comparing with other state-of-the-art fully convolutional networks. The transfer learning and dice loss function are beneficial to improve the performance of on the small and unbalanced dataset. Our best model on the dataset gets an overall accuracy of 82.9%, an average F1 score of 66.0% and an average IoU of 52.7%.