Mapping Rice Paddy Based on Machine Learning with Sentinel-2 Multi-Temporal Data: Model Comparison and Transferability

Paddy Rice mapping in fragmented lands by improved phenology curve and correlation measurements on Sentinel-2 imagery in Google earth engine

Accurate and timely rice crop mapping is important to address the challenges of food security, water management, disease transmission, and land use change. However, accurate rice crop mapping is difficult due to the presence of mixed pixels in small and fragmented rice fields as well as cloud cover. In this paper, a phenology-based method using Sentinel-2 time series images is presented to solve these problems. First, the improved rice phenology curve is extracted based on Normalized Difference Vegetation Index and Land Surface Water Index time series data of rice fields. Then, correlation was taken between rice phenology curve and time series data of each pixel. The correlation result of each pixel shows the similarity of its time series behavior with the proposed rice phenology curve. In the next step, the maximum correlation value and its occurrence time are used as the feature vectors of each pixel to classification. Since correlation measurement provides data with better separability than its input data, training the classifier can be done with fewer samples and the classification is more accurate. The implementation of the proposed correlation-based algorithm can be done in a parallel computing. All the processes were performed on the Google Earth Engine cloud platform on the time series images of the Sentinel 2. The implementations show the high accuracy of this method.

Karst vegetation coverage detection using UAV multispectral vegetation indices and machine learning algorithm

BACKGROUND

Karst vegetation is of great significance for ecological restoration in karst areas. Vegetation Indices (VIs) are mainly related to plant yield which is helpful to understand the status of ecological restoration in karst areas. Recently, karst vegetation surveys have gradually shifted from field surveys to remote sensing-based methods. Coupled with the machine learning methods, the Unmanned Aerial Vehicle (UAV) multispectral remote sensing data can effectively improve the detection accuracy of vegetation and extract the important spectrum features.

RESULTS

In this study, UAV multispectral image data at flight altitudes of 100 m, 200 m, and 400 m were collected to be applied for vegetation detection in a karst area. The resulting ground resolutions of the 100 m, 200 m, and 400 m data are 5.29, 10.58, and 21.16 cm/pixel, respectively. Four machine learning models, including Random Forest (RF), Support Vector Machine (SVM), Gradient Boosting Machine (GBM), and Deep Learning (DL), were compared to test the performance of vegetation coverage detection. 5 spectral values (Red, Green, Blue, NIR, Red edge) and 16 VIs were selected to perform variable importance analysis on the best detection models. The results show that the best model for each flight altitude has the highest accuracy in detecting its training data (over 90%), and the GBM model constructed based on all data at all flight altitudes yields the best detection performance covering all data, with an overall accuracy of 95.66%. The variables that were significantly correlated and not correlated with the best model were the Modified Soil Adjusted Vegetation Index (MSAVI) and the Modified Anthocyanin Content Index (MACI), respectively. Finally, the best model was used to invert the complete UAV images at different flight altitudes.

CONCLUSIONS

In general, the GBM_all model constructed based on UAV imaging with all flight altitudes was feasible to accurately detect karst vegetation coverage. The prediction models constructed based on data from different flight altitudes had a certain similarity in the distribution of vegetation index importance. Combined with the method of visual interpretation, the karst green vegetation predicted by the best model was in good agreement with the ground truth, and other land types including hay, rock, and soil were well predicted. This study provided a methodological reference for the detection of karst vegetation coverage in eastern China.

Multi-Crop Classification Using Feature Selection-Coupled Machine Learning Classifiers Based on Spectral, Textural and Environmental Features

Machine learning (ML) classifiers have been widely used in the field of crop classification. However, having inputs that include a large number of complex features increases not only the difficulty of data collection but also reduces the accuracy of the classifiers. Feature selection (FS), which can availably reduce the number of features by selecting and reserving the most essential features for crop classification, is a good tool to solve this problem effectively. Different FS methods, however, have dissimilar effects on various classifiers, so how to achieve the optimal combination of FS methods and classifiers to meet the needs of high-precision recognition of multiple crops remains an open question. This paper intends to address this problem by coupling the analysis of three FS methods and six classifiers. Spectral, textual, and environmental features are firstly extracted as potential classification indexes from time-series remote sensing images from France. Then, three FS methods are used to obtain feature subsets and combined with six classifiers for coupling analysis. On this basis, 18 multi-crop classification models (FS–ML models) are constructed. Additionally, six classifiers without FS are constructed for comparison. The training set and the validation set for these models are constructed by using the Kennard-Stone algorithm with 70% and 30% of the samples, respectively. The performance of the classification model is evaluated by Kappa, F1-score, accuracy, and other indicators. The results show that different FS methods have dissimilar effects on various models. The best FS–ML model is RFAA+-RF, and its Kappa coefficient can reach 0.7968, which is 0.33–46.67% higher than that of other classification models. The classification results are highly dependent on the original classification index sets. Hence, the reasonability of combining spectral, textural, and environmental indexes is verified by comparing them with the single feature index set. The results also show that the classification strategy combining spectral, textual, and environmental indexes can effectively improve the ability of crop recognition, and the Kappa coefficient is 9.06–65.52% higher than that of the single unscreened feature set.

A Dual Attention Convolutional Neural Network for Crop Classification Using Time-Series Sentinel-2 Imagery

Accurate and timely mapping of crop types and having reliable information about the cultivation pattern/area play a key role in various applications, including food security and sustainable agriculture management. Remote sensing (RS) has extensively been employed for crop type classification. However, accurate mapping of crop types and extents is still a challenge, especially using traditional machine learning methods. Therefore, in this study, a novel framework based on a deep convolutional neural network (CNN) and a dual attention module (DAM) and using Sentinel-2 time-series datasets was proposed to classify crops. A new DAM was implemented to extract informative deep features by taking advantage of both spectral and spatial characteristics of Sentinel-2 datasets. The spectral and spatial attention modules (AMs) were respectively applied to investigate the behavior of crops during the growing season and their neighborhood properties (e.g., textural characteristics and spatial relation to surrounding crops). The proposed network contained two streams: (1) convolution blocks for deep feature extraction and (2) several DAMs, which were employed after each convolution block. The first stream included three multi-scale residual convolution blocks, where the spectral attention blocks were mainly applied to extract deep spectral features. The second stream was built using four multi-scale convolution blocks with a spatial AM. In this study, over 200,000 samples from six different crop types (i.e., alfalfa, broad bean, wheat, barley, canola, and garden) and three non-crop classes (i.e., built-up, barren, and water) were collected to train and validate the proposed framework. The results demonstrated that the proposed method achieved high overall accuracy and a Kappa coefficient of 98.54% and 0.981, respectively. It also outperformed other state-of-the-art classification methods, including RF, XGBOOST, R-CNN, 2D-CNN, 3D-CNN, and CBAM, indicating its high potential to discriminate different crop types.

Rice Mapping in Training Sample Shortage Regions Using a Deep Semantic Segmentation Model Trained on Pseudo-Labels

A deep semantic segmentation model-based method can achieve state-of-the-art accuracy and high computational efficiency in large-scale crop mapping. However, the model cannot be widely used in actual large-scale crop mapping applications, mainly because the annotation of ground truth data for deep semantic segmentation model training is time-consuming. At the operational level, it is extremely difficult to obtain a large amount of ground reference data by photointerpretation for the model training. Consequently, in order to solve this problem, this study introduces a workflow that aims to extract rice distribution information in training sample shortage regions, using a deep semantic segmentation model (i.e., U-Net) trained on pseudo-labels. Based on the time series Sentinel-1 images, Cropland Data Layer (CDL) and U-Net model, the optimal multi-temporal datasets for rice mapping were summarized, using the global search method. Then, based on the optimal multi-temporal datasets, the proposed workflow (a combination of K-Means and random forest) was directly used to extract the rice-distribution information of Jiangsu (i.e., the K–RF pseudo-labels). For comparison, the optimal well-trained U-Net model acquired from Arkansas (i.e., the transfer model) was also transferred to Jiangsu to extract local rice-distribution information (i.e., the TF pseudo-labels). Finally, the pseudo-labels with high confidences generated from the two methods were further used to retrain the U-Net models, which were suitable for rice mapping in Jiangsu. For different rice planting pattern regions of Jiangsu, the final results showed that, compared with the U-Net model trained on the TF pseudo-labels, the rice area extraction errors of pseudo-labels could be further reduced by using the U-Net model trained on the K–RF pseudo-labels. In addition, compared with the existing rule-based rice mapping methods, he U-Net model trained on the K–RF pseudo-labels could robustly extract the spatial distribution information of rice. Generally, this study could provide new options for applying a deep semantic segmentation model to training sample shortage regions.

Machine Learning in Agriculture: A Comprehensive Updated Review

The digital transformation of agriculture has evolved various aspects of management into artificial intelligent systems for the sake of making value from the ever-increasing data originated from numerous sources. A subset of artificial intelligence, namely machine learning, has a considerable potential to handle numerous challenges in the establishment of knowledge-based farming systems. The present study aims at shedding light on machine learning in agriculture by thoroughly reviewing the recent scholarly literature based on keywords' combinations of "machine learning" along with "crop management", "water management", "soil management", and "livestock management", and in accordance with PRISMA guidelines. Only journal papers were considered eligible that were published within 2018-2020. The results indicated that this topic pertains to different disciplines that favour convergence research at the international level. Furthermore, crop management was observed to be at the centre of attention. A plethora of machine learning algorithms were used, with those belonging to Artificial Neural Networks being more efficient. In addition, maize and wheat as well as cattle and sheep were the most investigated crops and animals, respectively. Finally, a variety of sensors, attached on satellites and unmanned ground and aerial vehicles, have been utilized as a means of getting reliable input data for the data analyses. It is anticipated that this study will constitute a beneficial guide to all stakeholders towards enhancing awareness of the potential advantages of using machine learning in agriculture and contributing to a more systematic research on this topic.

Detection of Very Small Tree Plantations and Tree-Level Characterization Using Open-Access Remote-Sensing Databases

Highly fragmented land property hinders the planning and management of single species tree plantations. In such situations, acquiring information about the available resources is challenging. This study aims to propose a method to locate and characterize tree plantations in these cases. Galicia (Northwest of Spain) is an area where property is extremely divided into small parcels. European chestnut (Castanea sativa) plantations are an important source of income there; however, it is often difficult to obtain information about them due to their small size and scattered distribution. Therefore, we selected a Galician region with a high presence of chestnut plantations as a case study area in order to locate and characterize small plantations using open-access data. First, we detected the location of chestnut plantations applying a supervised classification for a combination of: Sentinel-2 images and the open-access low-density Light Detection and Ranging (LiDAR) point clouds, obtained from the untapped open-access LiDAR Spanish national database. Three classification algorithms were used: Random Forest (RF), Support Vector Machine (SVM), and XGBoost. We later characterized the plots at the tree-level using the LiDAR point-cloud. We detected individual trees and obtained their height applying a local maxima algorithm to a point-cloud-derived Canopy Height Model (CHM). We also calculated the crown surface of each tree by applying a method based on two-dimensional (2D) tree shape reconstruction and canopy segmentation to a projection of the LiDAR point cloud. Chestnut plantations were detected with an overall accuracy of 81.5%. Individual trees were identified with a detection rate of 96%. The coefficient of determination R2 value for tree height estimation was 0.83, while for the crown surface calculation it was 0.74. The accuracy achieved with these open-access databases makes the proposed procedure suitable for acquiring knowledge about the location and state of chestnut plantations as well as for monitoring their evolution.