Pixel-based spatiotemporal statistics from remotely sensed imagery improves spatial predictions and sampling strategies of alluvial soils

Abstract

Alluvial plains are vexing landscapes for soil mapping and spatial soil property predictions. Alluvial sediments often exhibit unpredictable spatial variability from both fluvial and anthropogenic disturbance. The determination of optimal number of soil sampling points for capturing soil variability remains a persistent issue for mapping and monitoring soil conditions. Here, soil organic matter (SOM) and cation exchange capacity (CEC) were estimated in an alluvial plain from a dense array of 2145 soil samples over 250 ha. The primary goals were to i) use pixel-based statistics from time-series Sentinel-2 satellite reflectance data to estimate SOM and CEC; ii) evaluate the use of images with vegetation cover versus bare soil images; and, iii) investigate the optimal number of sampling points to map alluvial soils using different sampling strategies. The optimal sample density was 1 sample per 2.5 ha based on the overlapping of prediction distributions (OV > 0.9). Conditioned Latin Hypercube Sampling (cLHS) was the most efficient sampling strategy. Random grid sampling provided the least consistent results. The use of cLHS coupled with the mean and standard deviation bands was the optimal sampling strategy. Pixel-based statistics from readily available satellite imagery captured persistent soil-reflectance relationships that enabled the use of row crops as proxies to predict soil properties. The combined use of pixel-based statistics as inputs to cLHS can reduce the chances of oversampling and associated costs. Pixel-based statistics provided persistent spatial information that can enhance precision agriculture management, carbon mapping, and is likely applicable to other soil properties and remotely sensed products. Use of pixel-based statistics from Sentinel-2 and similar global-coverage imagery could provide readily-calculable inputs for improved soil mapping and more optimal site sample selections in other agricultural row crop areas, globally.

Keywords

• Inceptisols
• Vertisols
• Soil organic carbon
• Cation exchange capacity
• Sentinel-2
• Machine learning