The number of samples used in the calibration data set affects the quality of the generated predictive models using visible, near and shortwave infrared (VIS–NIR–SWIR) spectroscopy for soil attributes. Recently, the convolutional neural network (CNN) has been regarded as a highly accurate model for predicting soil properties on a large database. However, it has not yet been ascertained how large the sample size should be for CNN model to be effective. This paper investigates the effect of the training sample size on the accuracy of deep learning and machine learning models. It aims at providing an estimate of how many calibration samples are needed to improve the model performance of soil properties predictions with CNN as compared to conventional machine learning models. In addition, this paper also looks at a way to interpret the CNN models, which are commonly labelled as a black box. It is hypothesised that the performance of machine learning models will increase with an increasing number of training samples, but it will plateau when it reaches a certain number, while the performance of CNN will keep improving. The performances of two machine learning models (partial least squares regression – PLSR; Cubist) are compared against the CNN model. A VIS–NIR–SWIR spectra library from Brazil, containing 4251 unique sites with averages of two to three samples per depth (a total of 12 044 samples), was divided into calibration (3188 sites) and validation (1063 sites) sets. A subset of the calibration data set was then created to represent a smaller calibration data set ranging from 125, 300, 500, 1000, 1500, 2000, 2500 and 2700 unique sites, which is equivalent to a sample size of approximately 350, 840, 1400, 2800, 4200, 5600, 7000 and 7650. All three models (PLSR, Cubist and CNN) were generated for each sample size of the unique sites for the prediction of five different soil properties, i.e. cation exchange capacity, organic carbon, sand, silt and clay content. These calibration subset sampling processes and modelling were repeated 10 times to provide a better representation of the model performances. Learning curves showed that the accuracy increased with an increasing number of training samples. At a lower number of samples (< 1000), PLSR and Cubist performed better than CNN. The performance of CNN outweighed the PLSR and Cubist model at a sample size of 1500 and 1800, respectively. It can be recommended that deep learning is most efficient for spectra modelling for sample sizes above 2000. The accuracy of the PLSR and Cubist model seems to reach a plateau above sample sizes of 4200 and 5000, respectively, while the accuracy of CNN has not plateaued. A sensitivity analysis of the CNN model demonstrated its ability to determine important wavelengths region that affected the predictions of various soil attributes.

The Earth’s surface dynamics provide essential information for guiding environmental and agricultural policies. Uncovered and unprotected surfaces experience several undesirable effects, which can affect soil ecosystem functions. We developed a technique to identify global bare surface areas and their dynamics based on multitemporal remote sensing images to aid the spatiotemporal evaluation of anthropic and natural phenomena. The bare Earth’s surface and its changes were recognized by Landsat image processing over a time range of 30 years using the Google Earth Engine platform. Two additional products were obtained with a similar technique: a) Earth’s bare surface frequency, which represents where and how many times a single pixel was detected as bare surface, based on Landsat series, and b) Earth’s bare soil tendency, which represents the tendency of bare surface to increase or decrease. This technique enabled the retrieval of bare surfaces on 32% of Earth’s total land area and on 95% of land when considering only agricultural areas. From a multitemporal perspective, the technique found a 2.8% increase in bare surfaces during the period on a global scale. However, the rate of soil exposure decreased by ~4.8% in the same period. The increase in bare surfaces shows that agricultural areas are increasing worldwide. The decreasing rate of soil exposure indicates that, unlike popular opinion, more soils have been covered due to the adoption of conservation agriculture practices, which may reduce soil degradation.

The mapping of soil attributes provides support to agricultural planning and land use monitoring, which consequently aids the improvement of soil quality and food production. Landsat 5 Thematic Mapper (TM) images are often used to estimate a given soil attribute (i.e., clay), but have the potential to model many other attributes, providing input for soil mapping applications. In this paper, we aim to evaluate a Bare Soil Composite Image (BSCI) from the state of São Paulo, Brazil, calculated from a multi-temporal dataset, and study its relationship with topsoil properties, such as soil class and geology. The method presented detects bare soil in satellite images in a time series of 16 years, based on Landsat 5 TM observations. The compilation derived a BSCI for the agricultural sites (242,000 hectare area) characterized by very complex geology. Soil properties were analyzed to calibrate prediction models using 740 soil samples (0–20 cm) collected of the area. Partial least squares regression (PLSR) based on the BSCI spectral dataset was performed to quantify soil attributes. The method identified that a single image represents 7 to 20% of bare soil while the compilation of the multi-temporal dataset increases to 53%. Clay content had the best soil attribute prediction estimates (R² = 0.75, root mean square error (RMSE) = 89.84 g kg⁻¹, and accuracy = 74%). Soil organic matter, cation exchange capacity and sandy soils also achieved moderate predictions. The BSCI demonstrates a strong relationship with legacy geological maps detecting variations in soils. From a single composite image, it was possible to use spectroscopy to evaluate several environmental parameters. This technique could greatly improve soil mapping and consequently aid several applications, such as land use planning, environmental monitoring, and prevention of land degradation, updating legacy surveys and digital soil mapping.

Soil spectroscopy is a widely used method for estimating soil properties that are important to environmental and agricultural monitoring. However, a bottleneck to its more widespread adoption is the need for establishing large reference datasets for training machine learning (ML) models, which are called soil spectral libraries (SSLs). Similarly, the prediction capacity of new samples is also subject to the number and diversity of soil types and conditions represented in the SSLs. To help bridge this gap and enable hundreds of stakeholders to collect more affordable soil data by leveraging a centralized open resource, the Soil Spectroscopy for Global Good initiative has created the Open Soil Spectral Library (OSSL). In this paper, we describe the procedures for collecting and harmonizing several SSLs that are incorporated into the OSSL, followed by exploratory analysis and predictive modeling. The results of 10-fold cross-validation with refitting show that, in general, mid-infrared (MIR)-based models are significantly more accurate than visible and near-infrared (VisNIR) or near-infrared (NIR) models. From independent model evaluation, we found that Cubist comes out as the best-performing ML algorithm for the calibration and delivery of reliable outputs (prediction uncertainty and representation flag). Although many soil properties are well predicted, total sulfur, extractable sodium, and electrical conductivity performed poorly in all spectral regions, with some other extractable nutrients and physical soil properties also performing poorly in one or two spectral regions (VisNIR or NIR). Hence, the use of predictive models based solely on spectral variations has limitations. This study also presents and discusses several other open resources that were developed from the OSSL, aspects of opening data, current limitations, and future development. With this genuinely open science project, we hope that OSSL becomes a driver of the soil spectroscopy community to accelerate the pace of scientific discovery and innovation.