Search

Spatial distributions of soil total nitrogen (STN) can greatly contribute to the precision fertilization and crop yield in black soil area. Many efforts have been devoted to the accurate algorithms for the estimation of STN contents. This study aims to firstly propose the applicable integrated machine learning algorithms (e.g., Random Forest (RF), Adaptive boosting (AdaBoost) and Gradient boosting categorical features (CatBoost)) and Supervised learning algorithms (e.g., Simple linear regression (SLR), Support vector regression (SVR) and Back propagation neural network (BPNN)). The spectral indexes and environmental variables were then integrated using Multispectral Imager (MSI) product, in order to seamlessly retrieve the spatial distributions of STN. A large number of soil samples were collected in Suihua City, and the synchronous reflectance that embedded in better quality of Sentinel-2 Level-2A images. Likewise, two scenarios were considered, e.g., band 1-12 reflectance or combining them with spectral indexes and environmental variables (digital elevation model, temperature, precipitation and soil types). The results showed that the average STN of in situ measured samples was 1 904.06 mg/kg, with a coefficient of variation of 17.93%. The coefficients of determination (R²) were smaller than 0.6 between the measured and derived values from the developed STN algorithms, when the band 1-12 reflectance as the input variables. The performances of six STN algorithms for the validated dataset were ranked in the descending order of RF, CatBoost, AdaBoost, BPNN, SLR and SVR, whereas, the importance were ranked in the order of RF, SVR, BPNN, AdaBoost, CatBoost, and SLR. Once the band 1- 12 reflectance, spectral indexes, and environmental variables were as the input variables, the performance of STN algorithms was improved significantly in the validated dataset, of which the R² increased by 0.22 and root mean square error (RMSE) decreased by 35.30 mg/kg. In total, the accuracies of STN algorithms were in the descending order of RF, CatBoost, AdaBoost, BPNN, MLSR, and SVR. Hence, the RF can be expected to simulate the nonlinear relationships between reflectance and STN, and then obtain a better degree of measured- and derived- fitting, indicating powerful nonlinear ability. Furthermore, the STN content was mapped using Sentinel-2 level2A imagery and RF algorithm, in order to examine the spatial variation. The spatial distribution of STN content was higher in the northeast, whereas lower in southwest-decreasing gradually from north to southand slightly higher in middle of Suihua City. This was attributed to the large number of environmental variables. Anyway, much more attention can be paid for the decision-making on the protection of ‘Black soil’ and natural ecosystems. The finding can provide the technical assistances on dynamically monitoring STN contents, in order to evaluate the soil fertility for the sustainable agricultural development in black soil area of Northeast China. [ABSTRACT FROM AUTHOR]

Background: Relying solely on radiata pine (Pinus radiata D.Don) leaves New Zealand's plantation forest industry vulnerable to fluctuations in market demand and at risk from a potentially devastating pest or disease outbreak. Therefore, the New Zealand government and forestry industry urge to diversify the forest resource and wood supply beyond the reliance on radiata pine. Unfortunately, the lack of accurate information on minor species' area, composition, and location poses challenges to forecasting potential log supply and logistics planning. Methods: The objective of this study is to classify minor species in New Zealand using imagery and phenological features extracted from data collected by the Copernicus Sentinel-2 satellite. The study collected reference data of minor species from large-scale forest owners and applied Random Forest classification using Sentinel-2 imagery to classify nine minor species classes in the Hawke's Bay region of New Zealand. Results: The study achieved an overall classification accuracy of 92.2% for minor species in New Zealand's Hawke's Bay region. Douglas-fir (Pseudotsuga menziesii (Mirb.) Franco) and Eucalyptus species had the highest accuracies, exceeding 90% for both producer's and user's accuracies. Acacia, larch, and other mixed species had lower accuracies, likely due to their lower occurrence. The most important input variable for classification was the Digital Elevation Model, indicating the significance of elevation in differentiating plantation species. The Greenness Index (GI) and Red edge bands also proved useful in the classification. The phenological measure Mean-EVI2 was found useful in classifying deciduous species such as larch and poplar. Conclusions: To the best of our knowledge, this study is the first to map the spatial extent and distribution of minor plantation species in New Zealand at the regional level, providing promising results for potentially expanding the study to national-level species mapping. [ABSTRACT FROM AUTHOR]

Land cover classification is able to reflect the potential natural and social process in urban development, providing vital information to stakeholders. Recent solutions on land cover classification are generally addressed by remotely sensed imagery and supervised classification methods. However, a high-performance classifier is desirable but challenging due to the existence of model hyperparameters. Conventional approaches generally rely on manual tuning, which is time-consuming and far from satisfying. Therefore, this work aims to propose a systematic method to automatically tune the hyperparameters by Bayesian parameter optimization for the random forest classifier. The recently launched Sentinel-2A/B satellites are drawn to provide the remote sensing imageries for land cover classification case study in Beijing, China, which have the best spectral/spatial resolutions among the freely available satellites. The improved random forest with Bayesian parameter optimization is compared against the support vector machine (SVM) and random forest (RF) with default hyperparameters by discriminating five land cover classes including building, tree, road, water, and crop field. Comparative experimental results show that the optimized RF classifier outperforms the conventional SVM and the RF with default hyperparameters in terms of accuracy, precision, and recall. The effects of band/feature number and the band usefulness are also assessed. It is envisaged that the improved classifier for Sentinel-2 satellite image processing can find a wide range of applications where high-resolution satellite imagery classification is applicable. [ABSTRACT FROM AUTHOR]

Snow coverage mapping plays a vital role not only in studying hydrology and climatology, but also in investigating crop disease overwintering for smart agriculture management. This work investigates snow coverage mapping by learning from Sentinel-2 satellite multispectral images via machine-learning methods. To this end, the largest dataset for snow coverage mapping (to our best knowledge) with three typical classes (snow, cloud and background) is first collected and labeled via the semi-automatic classification plugin in QGIS. Then, both random forest-based conventional machine learning and U-Net-based deep learning are applied to the semantic segmentation challenge in this work. The effects of various input band combinations are also investigated so that the most suitable one can be identified. Experimental results show that (1) both conventional machine-learning and advanced deep-learning methods significantly outperform the existing rule-based Sen2Cor product for snow mapping; (2) U-Net generally outperforms the random forest since both spectral and spatial information is incorporated in U-Net via convolution operations; (3) the best spectral band combination for U-Net is B2, B11, B4 and B9. It is concluded that a U-Net-based deep-learning classifier with four informative spectral bands is suitable for snow coverage mapping.

Land cover classification is able to reflect the potential natural and social process in urban development, providing vital information to stakeholders. Recent solutions on land cover classification are generally addressed by remotely sensed imagery and supervised classification methods. However, a high-performance classifier is desirable but challenging due to the existence of model hyperparameters. Conventional approaches generally rely on manual tuning, which is time-consuming and far from satisfying. Therefore, this work aims to propose a systematic method to automatically tune the hyperparameters by Bayesian parameter optimization for the random forest classifier. The recently launched Sentinel-2A/B satellites are drawn to provide the remote sensing imageries for land cover classification case study in Beijing, China, which have the best spectral/spatial resolutions among the freely available satellites. The improved random forest with Bayesian parameter optimization is compared against the support vector machine (SVM) and random forest (RF) with default hyperparameters by discriminating five land cover classes including building, tree, road, water, and crop field. Comparative experimental results show that the optimized RF classifier outperforms the conventional SVM and the RF with default hyperparameters in terms of accuracy, precision, and recall. The effects of band/feature number and the band usefulness are also assessed. It is envisaged that the improved classifier for Sentinel-2 satellite image processing can find a wide range of applications where high-resolution satellite imagery classification is applicable.

Snow coverage mapping plays a vital role not only in studying hydrology and climatology, but also in investigating crop disease overwintering for smart agriculture management. This work investigates snow coverage mapping by learning from Sentinel-2 satellite multispectral images via machine-learning methods. To this end, the largest dataset for snow coverage mapping (to our best knowledge) with three typical classes (snow, cloud and background) is first collected and labeled via the semi-automatic classification plugin in QGIS. Then, both random forest-based conventional machine learning and U-Net-based deep learning are applied to the semantic segmentation challenge in this work. The effects of various input band combinations are also investigated so that the most suitable one can be identified. Experimental results show that (1) both conventional machine-learning and advanced deep-learning methods significantly outperform the existing rule-based Sen2Cor product for snow mapping; (2) U-Net generally outperforms the random forest since both spectral and spatial information is incorporated in U-Net via convolution operations; (3) the best spectral band combination for U-Net is B2, B11, B4 and B9. It is concluded that a U-Net-based deep-learning classifier with four informative spectral bands is suitable for snow coverage mapping. [ABSTRACT FROM AUTHOR]