1-Balakrishna Gokaraju, 2-Sathish Samiappan, 1-Yogesh S. Kale and 1-Clinton Griffin
1-ViCAR Center and North Carolina A&T State University in NC, USA; 2-GRI and Mississippi State University in MS, USA
Multi-Sensor and Multi-Temporal Big Data Fusion using Deep Learning for UAS Precision Agriculture
Figure 1. Feature Extraction
Classification Problem :-- Multi-class Problem with 3 classes of Cotton, Corn and Soyabean.
Game Rules:-- DOWNLOAD Data from this Link :-- Click here
Participants will be provided with 100 tiles of training images each, for the above three classes of data. Total 300 tiles. (where each MSI tile is going to be size of 64*64 per tile) and each HSI tile is going to be 43*36 per same tile). Both tiles are of same ground spatial distance.
Participants will also be provided with “ground truth” output for these three classes of data. Folders-> class per Folder. Folder consists of Tiles of that class and date.
The 100 tiles for each class can be split in to training and validation. We removed the spatial reference (lat, long) of tiles to avoid hard coding.
Download Test Data from this Link :-- Click here
The Test Dataset will be a larger image comprising of mixed tiles of two classes and at different dates. Your model is going to be evaluated on the four test sites (testSite#1 to testSite#4). Your model can shall classify the three types of crops for level-1 classification and the three more temporal dates of the classes as early, mid and harvest ready crops for total 9 classes given in table 2 level-2 classification.
Your model performance is evaluated with each of your submission. Your submission will consist of the predicted class labels for 4 test sites in the form of classification map output. These map outputs will be of the same given test image matrix size. The evaluation scores will be Kappa accuracy against the ground truth class labels (not provided to you). Results and Top Scores will be announced shortly.
Please submit your classification map outputs to yskale@ncat.edu
Task :-- In precision agriculture sensitivity of multispectral and HSI imagery collected from time-series unmanned aerial vehicle (UAS), can be used to study and detect pesticide-induced stress at various scales in the carefully controlled experiments. The task is to train a Machine Learning (ML) or Deep Learning (DL) model for UAS Precision Agriculture Classification based on limited number of Image Tiles. The challenging task is the very high spatial resolution of HSI at 8cm and MS at 4 cm presents their own high definition radiometric resolution curse of speckle and noise. A Deep Learning model for Classification in UAS Precision Agriculture, might yield higher accuracy than an ML model.
Evaluation Criterion: --
Table 2. Class labels for Level-1 and Level-2 classification.
Crop Type
Class Labels for Level-1
Soyabean 1
Cotton 2
Corn 3
Crop Type for 9-class
Class Labels for Level-2
Early_Soyabean 4
Early_Cotton 5
Early_Corn 6
Mid_Soyabean 7
Mid_Cotton 8
Mid__Corn 9
Harvest_ready_Soyabean 10
Harvest_ready Mid_Cotton 11
Harvest_ready Mid__Corn 12
Level-1 Classification Kappa-Accuracy of the 3 class (Soyabean-1,Cotton-2 and Corn-3) Classification will be evaluated for 600 score points.
For Level-2 Classification, Additional 400 points for 9 more classes, per table 2. of the above three crops Classification (early crop, mid-crop, harvest ready-crop) using different month dates such as (June, July and September). Create Class labels using the table 2 for 9 class problem.
The top 10 Accuracy Scores will be assessed based on computation time and Kappa accuracy on the test data). Results and Top Scores will be announced on July 5th.
The computation times will be computed on ViCAR Machines and the criteria will be
HPC Score : ((120 - Computation time in sec)*0.62) + Kappa Accuracy *0.38
We provide the participants with guidelines regarding usage of our ViCAR-HPC resources.
Also, we provide the HPC resources for the game competition : Intel(R) Xeon(R) Gold 6148 CPU @ 2.40G 80 Cores, 2 NVidia GV100, GPU resources; 128 GB Ram; 100TB of storage. Reach out to Dr. Yogesh Kale at yskale@ncat.edu for computational resource account.
Special Issue Topic from this Competition: -- Comparison of Deep Learning and Machine Learning Techniques for Sensitivity Analyses of Crop Classification in Precision Agriculture.
Status of the Data :-
The dataset is original and has not been published yet.
Data is ready and also pre-processed, radiometrically calibrated.
The good ground truth is also available with class labels to the highest accuracy of GPS points (it is not provided to you for competition).
Acknowledgements and License of this Dataset is attributed to the Organizations “Visualizations and Computation Advancing Research (ViCAR) Center, Greensboro, NC; Geosystems Research Institute (GRI), Mississippi State, MS), and R. R. Foil Plant Science Research Center (RRFPSRC) at Mississippi State University.”
Special Issue Topic from this Competition: -- Comparison of Deep Learning and Machine Learning Techniques for Sensitivity Analyses of Crop Classification in Precision Agriculture.
Status of the Data :- -
The dataset is original and has not been published yet.
Data is ready and also pre-processed, radiometrically calibrated.
The good ground truth is also available with class labels to the highest accuracy of GPS points (it is not provided to you for competition).
Acknowledgements and License of this Dataset is attributed to the Organizations “Visualizations and Computation Advancing Research (ViCAR) Center, Greensboro, NC; Geosystems Research Institute (GRI), Mississippi State, MS), and R. R. Foil Plant Science Research Center (RRFPSRC) at Mississippi State University.”
The conference has agreed to support the competition with 200USD which will be awarded to the winning approach. Once competition is concluded and the winner is declared, we will request the necessary bank information from the winner.
