Key Research In Remote Sensing Image Collection And Processing Of Farmland

Comparing to the satellite-to-ground remote sensing and manned aircraft remote sensing, farmland information remote sensing using unmanned aerial vehicle (UAV) has obvious advantages, such as short revisit cycle, high resolution, low cost, high effiency and no space limit. In the future, it will be the primary way of remote sensing information collection in precision agriculture. However, UAV remote sensing in precision agriculture is still at the merging subject. With more and more attention being put in practical UAV remote sensing applications based on UAV, problems in practical application need to be addressed to promote the application of low-altitude remote sensing based on UAV, This paper put forward solutions for several problems existing in the low-altitude remote sensing based on UAV.(1) Before remote sensing data (image, spectrum, Lidar, SAR, etc.) used in farmland application, it needs to be georeferenced to specific ground point location. There is a lack of POS information recorder which fits for UAV remote sensing with small volume, light weight, low cost and good effect. This chapter describes the development of a POS data recorder which combines carrier phase RTK GNSS with MEMS inertial sensors in loosely couple navigation methods. It collects fast (200 Hz) inertial navigation data from MEMS BMX055 and slow RTK satellite navigation data(10Hz) from Swift Piksi GNSS receiver. STM32F405 was used as MCU to interprete data from Piksi using SBP and to communicate with BMX055 via ⅡC and store the collected onboard using self-defined minimum binary protocal. Ground data processing software was developed to translate binary txt into ASCII data and to generate continuous rapid positioning and attitude data. This software can help set recorder parameters like. Extended Kalman filter was used in integrated navigation algorithm to calculate the position and attitude of camera. Upper machine software was built for POS data processing and image georeferencing.(2) Static and dynamic test were done. In static test, recorder was placed at one point for 40 minutes. The recorded data was analysed in proposed method. the result shows that the precision of single GNSS receiver (CEP) was 0.187m, the processed positioning precision was 0.081m. In altitude processing result, confidential intervals of pitch, roll, yaw were 0.0031°、0.0032° and 0.0136° respectively. In dynamic test, the positioning result from integrated navigation algorithm shows better result than single GNSS.(3) Unmanned Aerial Vehicles (UAVs) have shown great potential in agriculture and are increasingly being developed for agricultural use. There are still a lot of experiments that need to be done to improve their performance and explore new uses, but experiments using UAVs are limited by many conditions like weather and location and the time it takes to prepare for a flight. To promote UAV remote sensing, a near ground remote sensing platform was developed. This platform consists of three major parts:(1) mechanical structures like a horizontal rail, vertical cylinder, and three axes gimbal; (2) power supply and control parts; (3) onboard application components. This platform covers five degrees of freedom (DOFs):horizontal, vertical, pitch, roll, yaw. A stm32 ARM single chip was used as the controller of the whole platform and another stm32 MCU was used to stabilize the gimbal. The gimbal stabilizer communicates with the main controller via a CAN bus. A multispectral camera was mounted on the gimbal. Software written in C++ language was developed as the graphical user interface. Operating parameters were set via this software and the working status was displayed in this software. Test was done to find out how this platform works. Laser distance meter was used to measure the slide rail’s repeat accuracy. A 3-axis vibration analyzer was used to test the system stability. Test results show that the horizontal repeat accuracy was less than 2 mm; vertical repeat accuracy was less than 1 mm; vibration was less than 2 g and remained at an acceptable level. The influence of platform vibration on remote sensing gimbal is verified using POS recorder and EKF methods. This system has high accuracy and stability and can therefore be used for various near ground remote sensing studies.(4) To integrate organic matte distribution information in 3d model, Structure from motion methods was used to generate plant’s 3D model. Tetracam ADC multispectral camera was used.31 multispectral images of a rape plant were collected at 3 different angles under indoor conditions for 3D reconstruction. Computer vision method-Structure from motion was used to process plant’s 3D model. The generated dense point cloud contains 120089 3D points.2682 points were removed as outliers. Control length from chess board is used to measure 3D model accuracy. The RMSE of spatial uniformity is 0.052599, and the maximum error is 0.1023cm. The result shows that this 3D model precisely represents rape plant’s morphology. To visualize the NDVI spatial distribution, a pseudo-color transform was performed according to color transformation theory. The result shows that the attempts to integrates multispectral image information into plant 3D reconstruction works out well and has a potential for plants’organic matters spatial distribution research.(5) Multi-source image fusion can reduce the miscalculation caused by using single source image. Using the complement and redundancies between multi-source data fusion between complete data can improve the reliability of data. This research was done based on the wavelet decomposition reconstruction method on near ground unmanned aerial vehicle (UAV) simulation platform. Two source image, respectively multispectral image and depth image of rapeseed were acquired. PMD camera, a kind of active imaging sensor which measure the time of flight of infrared light from generator to sensor was used to acquire depth image.3 kinds of image data was acquired from PMD camera namely distance image, intensity image and amplitude image. The 3 images were different forms of same data from sensor while the intensity image is more sensitive to the edge information in visual field. Tetracam ADC multispectral camera was used for multispectral image acquisition. The images acquired were calibrated using white board. To register the image, depth imaging principle was discussed and intensity image was used for camera internal parameter calibration. Pinhole model was used and chessboard calibration method was used for camera internal parameters. Since the distance image can’t see Harris points, intensity image was used for image calibration instead. The distance image and intensity image have same internal parameters. The internal parameters were final used for depth image correction so as to generate lens distortion-free images. SIFT method was used to find the corresponding points between 2 images. The first step is to find feature points descriptors, the second step is to use mahalanobis distance to find out proper matches.0.6 match ratio was found appropriate through tests for this purpose. After feature points were found and matched, multispectral image and depth image were resized to same zoom level then were registered and cropped to same size based on the corresponding points position. To find out which kind of wavelet basis fits best for this purpose, harr, Db2, Db4, Sym2, Sym4, Bior2.2, Bior2.4, Coif2, Coif4 was used on both source image for wavelet decomposition. Each decomposed wavelet layer was fused and formed another fused tower structure for wavelet reconstruction.5 parameters were used to evaluate fusion result, namely:Cross entropy, Average Gradient, Root Mean Square Error, Peak Signal-to-Noise Ratio and Mutual Information. The results shows that the Harr and sym4 wavelet basis have the best fusion result since they reserved most spectral information and texture information. To further analyze how wavelet decomposition level affect fusion result, images were decomposed to level 3,4,5 and 6 to find out which level fusion is better. The result was examined using the parameters to find out that the 3 and 4 fusion results were better than the 5 and 6. Then distance data in depth image was normalized to generate 3D point clouds. Statistical outlier removal method was used to remove outliers from point cloud. The final point cloud describes the spatial description of multispectral data. The research explored the fusion between multispectral image and depth image and studied the fusion parameters based on wavelet decomposition and reconstruction which would facilitate future use.