Integrating hyperspectral imaging with machine learning techniques for the high-resolution mapping of soil nitrogen fractions in soil profiles

Improved understanding of geographical variation and inequity in health status, wealth and access to resources within countries is increasingly being recognized as central to meeting development goals. Development and health indicators assessed at national or subnational scale can often conceal important inequities, with the rural poor often least well represented. The ability to target limited resources is fundamental, especially in an international context where funding for health and development comes under pressure. This has recently prompted the exploration of the potential of spatial interpolation methods based on geolocated clusters from national household survey data for the high-resolution mapping of features such as population age structures, vaccination coverage and access to sanitation. It remains unclear, however, how predictable these different factors are across different settings, variables and between demographic groups. Here we test the accuracy of spatial interpolation methods in producing gender-disaggregated high-resolution maps of the rates of literacy, stunting and the use of modern contraceptive methods from a combination of geolocated demographic and health surveys cluster data and geospatial covariates. Bayesian geostatistical and machine learning modelling methods were tested across four low-income countries and varying gridded environmental and socio-economic covariate datasets to build 1×1 km spatial resolution maps with uncertainty estimates. Results show the potential of the approach in producing high-resolution maps of key gender-disaggregated socio-economic indicators, with explained variance through cross-validation being as high as 74–75% for female literacy in Nigeria and Kenya, and in the 50–70% range for many other variables. However, substantial variations by both country and variable were seen, with many variables showing poor mapping accuracies in the range of 2–30% explained variance using both geostatistical and machine learning approaches. The analyses offer a robust basis for the construction of timely maps with levels of detail that support geographically stratified decision-making and the monitoring of progress towards development goals. However, the great variability in results between countries and variables highlights the challenges in applying these interpolation methods universally across multiple countries, and the importance of validation and quantifying uncertainty if this is undertaken.

Download Full-text

Experiments on Machine Learning Techniques for Soil Classification Using Sentinel-2 Products

10.5194/egusphere-egu2020-11080 ◽

2020 ◽

Author(s):

Victor Bacu ◽

Teodor Stefanut ◽

Dorian Gorgan

Keyword(s):

Machine Learning ◽

High Resolution ◽

Satellite Data ◽

Field Measurements ◽

Soil Classification ◽

Machine Learning Techniques ◽

Classification Models ◽

Spatial Features ◽

Learning Techniques ◽

Sentinel 2

Agricultural management relies on good, comprehensive and reliable information on the environment and, in particular, the characteristics of the soil. The soil composition, humidity and temperature can fluctuate over time, leading to migration of plant crops, changes in the schedule of agricultural work, and the treatment of soil by chemicals. Various techniques are used to monitor soil conditions and agricultural activities but most of them are based on field measurements. Satellite data opens up a wide range of solutions based on higher resolution images (i.e. spatial, spectral and temporal resolution). Due to this high resolution, satellite data requires powerful computing resources and complex algorithms. The need for up-to-date and high-resolution soil maps and direct access to this information in a versatile and convenient manner is essential for pedology and agriculture experts, farmers and soil monitoring organizations.Unfortunately, the satellite image processing and interpretation are very particular to each area, time and season, and must be calibrated by the real field measurements that are collected periodically. In order to obtain a fairly good accuracy of soil classification at a very high resolution, without using interpolation methods of an insufficient number of measurements, the prediction based on artificial intelligence techniques could be used. The use of machine learning techniques is still largely unexplored, and one of the major challenges is the scalability of the soil classification models toward three main directions: (a) adding new spatial features (i.e. satellite wavelength bands, geospatial parameters, spatial features); (b) scaling from local to global geographical areas; (c) temporal complementarity (i.e. build up the soil description by samples of satellite data acquired along the time, on spring, on summer, in another year, etc.).The presentation analysis some experiments and highlights the main issues on developing a soil classification model based on Sentinel-2 satellite data, machine learning techniques and high-performance computing infrastructures. The experiments concern mainly on the features and temporal scalability of the soil classification models. The research is carried out using the HORUS platform [1] and the HorusApp application [2], [3], which allows experts to scale the computation over cloud infrastructure.&#160;References:[1] Gorgan D., Rusu T., Bacu V., Stefanut T., Nandra N., &#8220;Soil Classification Techniques in Transylvania Area Based on Satellite Data&#8221;. World Soils 2019 Conference, 2 - 3 July 2019, ESA-ESRIN, Frascati, Italy (2019).[2] Bacu V., Stefanut T., Gorgan D., &#8220;Building soil classification maps using HorusApp and Sentinel-2 Products&#8221;, Proceedings of the Intelligent Computer Communication and Processing Conference &#8211; ICCP, in IEEE press (2019).[3] Bacu V., Stefanut T., Nandra N., Rusu T., Gorgan D., &#8220;Soil classification based on Sentinel-2 Products using HorusApp application&#8221;, Geophysical Research Abstracts, Vol. 21, EGU2019-15746, 2019, EGU General Assembly (2019).

Download Full-text

Identifying Mangrove Species Using Field Close-Range Snapshot Hyperspectral Imaging and Machine-Learning Techniques

Remote Sensing ◽

10.3390/rs10122047 ◽

2018 ◽

Vol 10 (12) ◽

pp. 2047 ◽

Cited By ~ 8

Author(s):

Jingjing Cao ◽

Kai Liu ◽

Yuanhui Zhu ◽

Jun Li ◽

Zhi He

Keyword(s):

Machine Learning ◽

Discriminant Analysis ◽

Hyperspectral Imaging ◽

Reflectance Spectra ◽

Machine Learning Techniques ◽

Order Derivative ◽

Mangrove Species ◽

First Order ◽

Learning Techniques ◽

Close Range

Investigating mangrove species composition is a basic and important topic in wetland management and conservation. This study aims to explore the potential of close-range hyperspectral imaging with a snapshot hyperspectral sensor for identifying mangrove species under field conditions. Specifically, we assessed the data pre-processing and transformation, waveband selection and machine-learning techniques to develop an optimal classification scheme for eight mangrove species in Qi'ao Island of Zhuhai, Guangdong, China. After data pre-processing and transformation, five spectral datasets, which included the reflectance spectra R and its first-order derivative d(R), the logarithm of the reflectance spectra log(R) and its first-order derivative d[log(R)], and hyperspectral vegetation indices (VIs), were used as the input data for each classifier. Consequently, three waveband selection methods, including the stepwise discriminant analysis (SDA), correlation-based feature selection (CFS), and successive projections algorithm (SPA) were used to reduce dimensionality and select the effective wavebands for identifying mangrove species. Furthermore, we evaluated the performance of mangrove species classification using four classifiers, including linear discriminant analysis (LDA), k-nearest neighbor (KNN), random forest (RF), and support vector machine (SVM). Application of the four considered classifiers on the reflectance spectra of all wavebands yielded overall classification accuracies of the eight mangrove species higher than 80%, with SVM having the highest accuracy of 93.54% (Kappa=0.9256). Using the selected wavebands derived from SPA, the accuracy of SVM reached 93.13% (Kappa=0.9208). The addition of hyperspectral VIs and d[log(R)] spectral datasets further improves the accuracies to 93.54% (Kappa=0.9253) and 96.46% (Kappa=0.9591), respectively. These results suggest that it is highly effective to apply field close-range snapshot hyperspectral images and machine-learning classifiers to classify mangrove species.

Download Full-text