Detecting Soil Moisture Content Based on Multispectral Images from Unmanned aerial vehicle

• Published in: 2023 Asia Symposium on Image Processing (ASIP) pp. 32 - 42
• Main Authors: Jinyu, Hao, Tianyou, Jiang, Shutian, He, Xinmeng, Liu, Jiming, Hu, Xiaoyong, Sun
• Format: Conference Proceeding
• Language: English
• Published: IEEE 15.06.2023
• Subjects: Autonomous aerial vehicles; Crops; Data models; Machine Learning; Machine learning algorithms; Multispectral; Soil moisture; Soil Moisture Content; Training; UAV; Vegetation mapping
• DOI: 10.1109/ASIP58895.2023.00014

The development of agriculture has been constrained by water shortage, and how to use new technology to detect the soil moisture content condition of farmland in real time is the key to achieve rational planning of agricultural water use. In this study, we use unmanned aerial vehicle (UAV) to collect canopy multispectral images of corn, cotton and peanut crops, extract spectral reflectance and calculate vegetation indices to construct data sets respectively. Using machine learning algorithms to establish a soil moisture content detection model based on UAV multispectral images to achieve rapid and accurate detection of soil moisture content in agricultural fields. The study set up a reasonable data acquisition scheme for multispectral image acquisition by considering the planting areas of corn, cotton and peanut in the experimental field, soil moisture conditions and flight conditions of the UAV. Using eight parameters as independent variables and soil moisture content as dependent variables, the data sets of the three crops were constructed separately and the training and test sets were divided in the ratio of 7:3. Random Forest (RF), Support Vector Regression (SVR), Backpropagation Neural Network(BPNN), Gradient Boosting Decision Tree (GBDT), bagging, and XGBoostmodels were built for soil moisture content detection, respectively. 1According to the model evaluation indexes, the Bagging model performed well in the corn multispectral soil moisture content dataset, with R 2 , RMSE and RPD reaching 0.809, 0.059 and 2.294, respectively, which can detect soil moisture content more accurately than other models. 2 To investigate whether the soil moisture content detection model established using the corn dataset can be extended to other crops, the In order to investigate whether the soil moisture content detection model developed using the corn dataset can be extended to other crops, the datasets based on multispectral images of cotton and peanut were brought into the model for testing. The results showed that the detection effect of peanut was better than that of cotton, but there was still a big difference between the detection effect of peanut and that of corn. 3The soil moisture content detection models based on multispectral data of cotton and peanut were established separately, and the best detection model for both crops was the GBDT model. The R 2 , RMSE and RPD of the best detection model for cotton soil moisture content reached 0.752, 0.065 and 2.009, respectively, while the R 2 , RMSE and RPD of the best detection model for peanut soil moisture content reached 0.738, 0.061 and 1.954, respectively, with high detection accuracy. Based on machine learning algorithms, this study builds six models to analyze the multispectral image data acquired by UAV, and also explores the effect of soil moisture content detection for various crops, in order to provide new strategies for future soil moisture content detection methods.