Yes. You can try deep learning for classification tasks.

Hey there,

Yes, Deep Learning and Machine Learning can be useful for that. I think you can find solutions (similar works) at http://www.patreo.dcc.ufmg.br/

Artur has rightly given the link.

Dear Satej,

This is a very complex and unsolved problem. Crop mapping in my experience is the most complicated issue in remote sensing due to many factors dynamic use of lands, size of crop fields, climatic conditions, and farmers general practices. Add to these factors lack of cheap technologies which can help in solving easily. Most of the time hyperspectral images with high spectral resolution is needed which can only be obtained using costly aerial surveys and pre collection of detailed spectral signatures of the crops during the phenological stages. I experienced this issue when I collected spectral signatures for winter wheat and alfalfa. There is no way to distinguish between these crops except in the case of high resolution spatial and spectral hyperspectral images. Of course one can use machine learning if known conditions exists such as planted day of the crop, knowledge of the phenological cycle (planting to harvesting date), and good algorithm. However, these knowledge are hindered by many factors such as the differences in the time of plantation between different farmers (i.e. some plant potato in early March others in early April), some crop cycles are similar, and the pre-requisite for good algorithm to succeed is to provide reliable data.

Please check the following projects and publications for more information about crop mapping.

Preprint CSSIT a new crop spectral signatures interactive tool

Conference Paper Possible field crops mapping and assessment in Lebanon using...

@M. Awad.. Thanks for your answer. Actually, extremely high spatial resolution makes it easier to extract the correct crop signatures. But, we can't get very high temporal resolution along with extremely high GSD. Planet is the only imagery which provides us 3m resolution as well as high repeativity. This is important in case of crop monitoring as we need to process images per week. In case of Hyperspectral, we don't have a satellite as of now, which gives images with very high temporal resolution. There will always be a trade-off between spatial, spectral and temporal resolution.

Sure I agree with you Satej, so that is why there is no one solution for this complex problem and I am working on it by using spectral signatures during the crop growth stages to identify the crops easily and to improve the crop mapping. As you can see in one paper I combined temporal resolution and knowledge of crop growth stage to identify crops and this can be automated using machine learning algorithm.

yes deep learning can perform well. please check if the data involve high frequency information then ResNet or DenseNet will perform better.

Hi,

I worked in the last two years on a similar project. My experience was, that the combination of different machine/statistical learning tools (deep CNN was one of them) worked very well. I used different feature extraction and classification methods and combined them using evidence pooling.

Your classification is based on hyperspectral images?

I analysed multispectral images some years ago for plant classification. In my case, the dataset was very well separable using simple classification approaches.

Wilfried Wöber . My classification is based on Planet imagery (https://www.planet.com/products/planet-imagery/). It is 3m multispectral imagery having very good temporal resolution. But, in my case, I am unable to separate different crop signatures.

Hello Satej,

unfortunately, without data I cannot suggest any algorithm. I had a student, which was able to classify plants based on multispectral images using a SVM.

In my opinion you have two choices:

- Manual Feature extraction (PCA, ICA, SGPLVM, etc.) and classification (SVM, kNN, random forest, ...)

- Deep Learning: CNN's (check out https://adeshpande3.github.io/The-9-Deep-Learning-Papers-You-Need-To-Know-About.html for initial understanding). I would not recommend you other deep architectures like deep Gaussian Processes.

In my opinion, both approaches have advantages.

Wilfried Wöber thanks .. The Analysis becomes complex since we have to deal with different phenological changes among crops. If a type of crop is sown in a particular area, then at that time, the other crop in that area may have reached its vegetative stage. So, the signatures of these two crops are sometimes similar which in turn reduce the spectral separability. This affects the accuracy of classification.

Hello Satej,

if I get u right - you will have plant state specific classification? For example young wheat have different spectral data than full grown wheat. Have you tried to cluster the data?

Have you tried to use statistical tools (e.g.: PCA)?

Wilfried Wöber . I am working on district wise crops as well as sub-district wise crops. In a particular sub-district, suppose there are two or three crops in a plot (having different sowing periods). I am supposed to classify images of this area per week. In a week's time, the signatures of different crops look similar. I tried using cluster (unsupervised) analysis, but still, accuracy is not improving. We are working on only 4 bands (RGB and NIR). I want to classify crop-wise not stage-wise. Different crops are there which I need to classify, I am not concentrating on a particular time, but every image showing different crops (with different growing stages of different crops) .

So you try to classify different crops using a single model which was trained using the whole dataset (different classes and different stages)?

Wilfried Wöber. I am trying to find a suitable algorithm to classify the images. Our previous algorithms produced low accuracy outcomes.

Are you using feature extraction/selection before classification? This would improve the results.

I would suggest PCA and SVM as a initial step... (e.g.: [1])

Of course, the "simples" way (from a development perspective) would be using a deep CNN for your classification...

[1] Kang, X. ; Xiang, X.; Li, L. and Benediktsson, J.A.: PCA-Based Edge-Preserving Features for Hyperspectral Image Classification, IEEE TRANSACTIONS ON GEOSCIENCE AND REMOTE SENSING, VOL. 55, NO. 12, DECEMBER 2017

Wilfried Wöber . Thanks for the valuable suggestions.

Dear Satej , maybe the crops classification based on feature fusion of hyperspectral and multispectral images using CNN (densenet is very useful in this time in computer vision but i think no one try it yet in remote sensing) can give you good results but it is really important to find the hyperspectral and multispectral images with high spectral , spatial and temporal resolution for the same study area of corse. This study it is very hard and complex specificly for the different creterion of the temporel resolution and it is really difficult to find the hyperspectral and multispectral data with these high resolution and also with their ground truth data for validation.

i don't focus on this classification in this time but i read some documents about it and i in my opinion Professor M.Awad gave you good advices and explications.

Best regards and good luck.

Dear sir

With due respect!

I think phenological metrics will be helpful for such complex classification using decision tree (DT) classification technique.

Avinash Kumar Ranjan.. Would you please elaborate what do you mean by phenological "metrics" and how these metrics would lead to good classification using DT?

Last year I have completed a project regarding Wheat acreage estimation for Sahibganj district. Wherein I had used time series datasets to derive time-series NDVI corresponding to Rabi season. Later, 10 GPS points of individual crop i.e. Wheat,mustard, and other pulses, etc.were used to extract the NDVI value with respect to time. Finally , a phenological graph were plotted to understand the variation in NDVI value of specific crop in different time. Based on the hypothesis, NDVI range of specific crop were used to make a DT rules for classification.

With respect

Avinash

Phenological graph means NDVI (an indicator of phenology??) vs Time, right? Which hypothesis did you use?

Yes... the DT rules can be derived based on the interpretation of NDVI ranges lies for specific crop in particular time period.

I could send you the article but now it is under review so I can't.

Yes. You can try deep learning for classification tasks.

I hope this link will help you

https://pdfs.semanticscholar.org/presentation/b74b/99303612b72489aea3ddaa1bdaf96cdaa01d.pdf

regards

Nice notes