Erratic rainfall with varying intensity and duration has raised the risks of crop failure in semi-arid areas of south and south-east Asia. In subsistence irrigation cropping systems often it’s difficult to schedule the irrigation, i.e. when and how much water to irrigate. Therefore there is a need for a regional real / near real-time updated database on vegetation greening and browning to facilitate the irrigation scheduling decisions. With the advent of open archives of remote sensing from United States Geological Survey (USGS) and European Space Agency (ESA) have proven a unique set of long-term historical and near real-time observations. In this study, an attempt has been made to understand the vegetation greening and browning patterns using time series of remote sensing observations for irrigation water management. The main objective is to study the greening and browning of natural vegetation (i.e., grasslands and forests) and agricultural areas of Indian sub-continent for understanding the breaks in the rainfall spells and integrated approach for irrigation scheduling. The time series of vegetation indices have been extracted for predefined grid locations from Sentinel 2 remote sensing sensor. Further, an algorithm based on time series analysis were evaluated for estimating the vegetation growth stages. The estimated vegetation growth stages was compared with the agro-climatic zones. A methodology for subsistence irrigation scheduling has been proposed based on regional vegetation growth stages (i.e. onset, peak and end of the season). The estimated vegetation growth stages showed poor alignment with the agro-climatic zones. The integrated approach based on vegetation growth stages is promising for scheduling subsistence irrigation. The proposed methodology for vegetation growth stage identification has potential applications in drought risk assessment and in establishing key indicators for agro-climatic zones.
Accurate and reliable information on spatio-temporal extent of surface water is critical for various agriculture/environmental applications such as drought, flood monitoring, and understanding the availability of surface water for irrigation. Remote sensing (Optical as well as SAR) datasets are extremely useful to monitor sur- face water at massive scale. In monsoon months the optical remote sensing observations over semi-arid Indian sub-continent are obstructed due to cloud cover. Synthetic Aperture Radar (SAR) is a useful alternative for year-round monitoring of the surface water bodies. Sentinel-1A and 1B are very useful to monitor the changes at very high spatial resolution and frequently due to its high spatiotemporal resolution. The main objective is to establish an operational methodology for estimation of spatiotemporal variations in the surface water availability using Sentinel-1A and 1B observations. The study has been carried out in four districts of Coastal Andhra Pradesh, India viz. Guntur, Krishna, East Godavari, and West Godavari. Training data for water vs. non-water (vegetation, forest, settlements, and barren lands) classes have been obtained from field visits and high-resolution Google Map overlay in Google Earth Engine. We divided the dataset into 70% data for model training and 30% for validation and evaluated the performance of tuned random forest classifier on the validation dataset. Results show the classification accuracy of 94.32%. Further, current and historical weather observations such as rainfall were used to assess the validity of spatiotemporal surface water layers. We found a good agreement between the rainfall and surface water availability. We observed the increase in the surface water area during July-August months due to rainfall as well as flooding in the rice fields during transplanting. We propose to use the crop area map, spatiotemporal surface water layers and weather observations for drought assessment i.e., historical drought events and areas prone to agricultural drought.
Timely and accurate recognition of health conditions in crops helps to perform necessary treatment for the plants. Automatically localizing these conditions in an image helps in estimating their spread and severity, thus saving on precious resources. Automated disease detection involving recognition as well as localization helps in identifying multiple diseases from one image and can be a small step forward for robotic farm surveying and spraying. Recent developments in Deep Neural Networks have drastically improved the localization and identification accuracy of objects. We leverage the neural network based method to perform accurate and fast detection of the diseases and pests in tea leaves. With a goal to identify an accurate yet efficient detector in terms of speed and memory, we evaluate various feature extraction networks and detection architectures. The images used to train and evaluate the models are with different resolutions, quality, brightness and focus as they are captured with mobile phones having different cameras through a participatory sensing approach. The experimental results show that the detection system effectively identifies and locates the health condition on the tea leaves in a complex background and with occlusion. We have evaluated YOLO based detection methods with different feature extraction architectures. Detection using YOLOv3 achieves mAP of about 86% with 50% IOU while making the system usable in real time.
Various pests and diseases can deteriorate the quality and yield of the capsicum. In order to control these losses, their timely detection is important. Thrips is one of the major pests in capsicum which is unable to detect in initial phase as the symptoms are not visible to naked eyes. Thrips not only causes plant damage but for the serious plant diseases it vectors. In this paper, we address the problem of detection of low infestation of thrips on capsicum leaves using multi-temporal hyperspectral remote sensing data simulated to multispectral sensors such as Sentinel-2 and Tetracam RGB+3. The reectance data from capsicum leaves with healthy and low infestations of thrips has been collected using handheld spectroradiometer. The hyperspectral remote sensing data is collected from 213 bands with wavelength ranging from 350 nm to 1052 nm and bandwidth varying from 3.22 nm to 3.346 nm during the period of 17 Mar to 13 Apr 2017. Variations observed in the spectral reflectance over time makes the detection based on multi-temporal data difficult. We have evaluated the performance of tuned random forest classifier for various set of features such as full feature set of 213 bands, features selected by Least Absolute Shrinkage and Selection Operator (LASSO) from 213 bands, features simulated to broad bands similar to Sentinel 2 and features simulated to multispectral bands similar to Tetracam RGB+3 (a camera which can be placed on drones). Results suggests that an overall classification accuracy of 92.81 % has been achieved on validation dataset using full feature set whereas accuracy slightly dips down to 90.3, 85.13 and 87.45 % when using selected features by LASSO, bands simulated to Sentinel-2 and Tetracam-RGB+3 respectively. Results imply that, Tetracam-RGB+3 and Sentinel-2 satellite can be effectively used for detection of low-infestation of thrips on capsicum.
Pesticide residues in the fruits, vegetables and agricultural commodities are harmful to humans and are becoming a health concern nowadays. Detection of pesticide residues on various commodities in an open environment is a challenging task. Hyperspectral sensing is one of the recent technologies used to detect the pesticide residues. This paper addresses the problem of detection of pesticide residues of Cyantraniliprole on grapes in open fields using multi temporal hyperspectral remote sensing data. The re ectance data of 686 samples of grapes with no, single and double dose application of Cyantraniliprole has been collected by handheld spectroradiometer (MS- 720) with a wavelength ranging from 350 nm to 1052 nm. The data collection was carried out over a large feature set of 213 spectral bands during the period of March to May 2015. This large feature set may cause model over-fitting problem as well as increase the computational time, so in order to get the most relevant features, various feature selection techniques viz Principle Component Analysis (PCA), LASSO and Elastic Net regularization have been used. Using this selected features, we evaluate the performance of various classifiers such as Artificial Neural Networks (ANN), Support Vector Machine (SVM), Random Forest (RF) and Extreme Gradient Boosting (XGBoost) to classify the grape sample with no, single or double application of Cyantraniliprole. The key finding of this paper is; most of the features selected by the LASSO varies between 350-373nm and 940-990nm consistently for all days. Experimental results also shows that, by using the relevant features selected by LASSO, SVM performs better with average prediction accuracy of 91.98 % among all classifiers, for all days.
Access to the requested content is limited to institutions that have purchased or subscribe to SPIE eBooks.
You are receiving this notice because your organization may not have SPIE eBooks access.*
*Shibboleth/Open Athens users─please
sign in
to access your institution's subscriptions.
To obtain this item, you may purchase the complete book in print or electronic format on
SPIE.org.
INSTITUTIONAL Select your institution to access the SPIE Digital Library.
PERSONAL Sign in with your SPIE account to access your personal subscriptions or to use specific features such as save to my library, sign up for alerts, save searches, etc.