Mapping Rice Paddy Based On Machine Learning With Sentinel-2Multi-Temporal Data

As the food crop with the largest planting area and the largest total production in China,rice has always been highly valued by governments and public.It is also an important basis for national and regional food policy and economic development plans.Therefore,accurately obtaining rice production information and accurately and timely grasping the status of crop production have very important practical significance for food security.According to statistics,about a quarter to a third of the world’s freshwater resources are used for rice paddy irrigation.In addition,rice paddy have been identified as important sources of methane（CH₄）and have a significant impact on the global greenhouse effect.Therefore,it is essential to map the spatial distribution and planting area of paddy rice at large scale for guiding rice production,water utilization,climate change,and government policy decisions.Rice agriculture is an important part of grain planting in the Southwest Hilly Area,China,but there has been a lack of efficient and accurate monitoring methods in the region due to the impact of factors such as field fragmentation,differences in farming systems,complex terrain,and remote sensing data quality.Recently,Convolutional Neural Networks（CNNs）have obtained considerable achievements in the remote sensing community.However,it has not been widely used in mapping rice paddy,and most studies lack the comparison of classification effectiveness and efficiency between CNNs and other classic machine learning models and their transferability.This study aims to develop various machine learning classification models with remote sensing data for comparing local accuracy of classifiers and evaluating transferability of pre-trained classifiers.Therefore,two types of experiments were designed:local classification experiments and model transferability experiments.These experiments were conducted using cloud-free Sentinel-2 multi-temporal data in Banan District and Zhongxian County,typical hilly areas of southwestern China.A pure pixel extraction algorithm was designed based on land-use vector data and Google Earth Online image.Four convolutional neural network（CNN）algorithms（one-dimensional（Conv1D）,two-dimensional（Conv2D）and three-dimensional（Conv3D＿1 and Conv3D＿2）convolutional neural networks）were developed and compared with four widely used classifiers（Random Forest（RF）,eXtreme Gradient Boosting（XGBoost）,Support Vector Machine（SVM）and Multi-Layer perceptron（MLP））.Recall,precision,Overall Accuracy（OA）and F1 score were applied to evaluate classification accuracy.The main conclusions are as follows:（1）The F1 scores of all classification models in the full-time combination experiment（Apr&Jul&Oct）are higher than the corresponding classifications in the single-phase feature experiment（Apr,Jul,and Oct）and the dual-phase combination experiment（Jul&Oct,Apr&Jul,Apr&Oct）model.It can be seen that the optimal temporal feature combination extracted from the spatial distribution of rice in the study area is a full temporal feature combination.In addition,in the single-phase feature experiment,all classifiers in Apr have the lowest F1 score;in the two-phase combination experiment,the classification model in Jul&Oct is better than Apr&Jul and Apr&Oct.（2）In local classification experiments,Conv2D performed best in local classification experiments with OA of 95.60%and F1 score of 0.9086 in Banan District,OA of 96.56%and F1 score of 0.9306 in Zhongxian County.The Conv-1D achieved the lowest classification accuracy,with F1 scores of 0.8048 and 0.7415,and overall accuracy of 89.87%and 83.42%,respectively.Conv2D performed best in local classification experiments with OA of 95.60%and F1 score of 0.9086 in Banan District,OA of 96.56%and F1 score of 0.9306 in Zhongxian County.（3）In model transferability experiments,almost all CNNs classifiers had low transferability.RF and XGBoost models have achieved acceptable F1 scores for transfer（RF=0.6799 and 0.6382,XGBoost=0.7187 and 0.6890,respectively）.Obviously,the transferability of XGBoost is better than RF.Moreover,compared with local classification experiments,the transfer classification accuracy of RF and XGBoost is lower than the local classification accuracy.（4）The results showed that Conv-2D with massive labeled samples coud obtain excellent classification accuracy in rice paddy mapping,while with non-labeled samples,pre-trained XGBoost provides higher classification performance.（5）The optimal classification model（Conv-2D in local classification experiments）was used to map rice paddy in Zhongxian County and Banan District.The predicted areas were 442,000 mu and 351,000 mu,with accuracy of 96.38%and 95.44%,respectively.