Search

• Onboard plant counting framework for efficient early crop emergence estimation. • Direct estimation of plant counts from aerial images, eliminating the need for costly orthomosaic formation. • Evaluation of three methods based on time complexity and accuracy, with average F1 scores ranging from 0.947 to 0.967. Crop emergence estimation at early crop growth stages is becoming increasingly important for the long-term sustainability of natural resources. It helps farmers and agricultural stakeholders in the efficient allocation of resources like water, pesticides, and fertilizers. It can be used to estimate the yield and seed quality, identify the region of potential yield losses, and make future agriculture plans. These future agriculture plans can play a crucial role in ensuring maximum crop population and yield while utilizing the same limited land and natural resources. Most of the existing plant counting frameworks require offline processing of images with computationally expensive algorithms including the structure for motion and multiview stereo to develop an orthomosaic. This study proposed a tobacco plant counting framework that directly estimates counts from aerial images and has the potential for real-time applicability. It consists of three core modules: overlap detection, plant detection, and plant counting. The overlap detection module replaces the need for computationally expensive orthomosaic formation to avoid counting repetition by overlap masking based on only visual cues. Three different methods are evaluated as core modules for finding an optimal solution for plant counting based on time complexity and accuracy. In the first method after overlap detection, semantic segmentation with U-NET is employed as a plant detection module. For plant counting, we count the connected pixels classified as plants to estimate the crop count. In the second method after overlap detection, object detection using YOLOv7 is utilized as a plant detection module followed by simply counting each detected plant. In the third method, we utilize YOLOv7 for object detection, similar to the second method. However, we introduce the SORT (Simple Online and Realtime Tracking) algorithm for object tracking. This object tracking replaces the overlap detection module making it a real-time applicable method. For plant counting, we assess the number of tracked plants. The proposed algorithm is evaluated on two distinct tobacco fields. The high-resolution aerial data is collected from tobacco fields near Peshawar, Pakistan, and is human-labelled. The first and second methods show average F1 scores of 0.947 and 0.9667, respectively, whereas the third method has the potential for real-time applicability with an average F1 score of 0.967. [ABSTRACT FROM AUTHOR]

• A new method for rice variety identification. • The classifier takes both spectral and time series information of each pixel. • Spectral unmixing model is used for labelling each pixel. • Radiometric indices are used to increase classification accuracy. • Experimental results have exhibited an overall accuracy of 98.6%. In the agriculture sector food productivity, security, and sustainability, imposed challenges on farmers, regulatory bodies, and policymakers due to increasing demand and depleting natural resources and environmental concerns. Rice crop holds a prominent place in Pakistan's agriculture sector, it is not only consumed locally but also exported to many countries including China. Gathering crop information such as variety maps, yield estimation, or etc. can help farmers, regulatory bodies, policymakers, and rice mills in decision-making. In Pakistan, crop information is collected through manual field surveys that require a lot of human labor, are costly, and are time-consuming. One cannot ignore human error and bias in the process. A new framework for pixel classification is proposed that uses both spectral and time-series data of Sentinal-2 satellite for mapping two rice varieties "Basmati" and "IRRI, grown in Pakistan. The data were collected from twelve rice fields (approx. 307 acres) of different geographical locations at 16-time instances to cover the complete rice-growing season (May–October) in 2019. A linear spectral unmixing model is used to determine sub-pixel information of water, soil, and vegetation content, which is used for labeling each pixel for supervised learning. The input to our classifier is a 16 × 15 image formed using 15 spectral features (12 spectral bands and 3 radiometric indices) of 16 carefully selected different time instances for each pixel. The output is a pixel-level classification (semantic segmentation) of each pixel into Basmati, IRRI, and others (soil, water, etc.). Experimental results have exhibited an excellent overall accuracy of 98.6% with the proposed approach. The Basmati rice obtained higher accuracy of 99.7% as compared to IRRI rice with an accuracy of 95.2%. [ABSTRACT FROM AUTHOR]