I presume that you're doing a supervised classification.
Usually, when you manually assign classes to points (cells) on the scene, ratio between training and testing samples should be 70% : 30%.
For accuracy assessment, it's desired to have at least 500 points total, from which the smallest classes should have at least 50 samples and bigger classes proportionally more samples.
If you want to satisfy the above conditions, you need ~ 1200 more samples (cells) for training.
That is 1700 points in total, which is pretty much.
Sampling based on single cells (points) is time-consuming. I would recommend you to select sampling polygons around randomly selected points (if they represent the same class) rather than sampling single cells. Doing this, you'll increase a number of training samples in shorter period.
You need to consider the size (area) of the region you are studying along with how many classes of land use / land cover are present or you are going to identify. Based on this you can do the sampling points. It is always advisable that, you need to take more number of points for classification than the accuracy assessment.
I hope this will help to think about the strategy you need to make while planning for the classification.
If you have more points, the result will be better (accuracy), depends only of his work and of the characteristics of the study area (total area, number of classes, image, ....)
Taking into consideration what has being said by others, and based on my personal experience, the more points you sample, the higher your accuracy.
If you can lay your hands on this papers (below), they will give you a better understanding on selecting sample points for accuracy assessment:
Foody, G. M. (2002). Status of land cover classification accuracy assessment. Remote sensing of environment, 80(1), 185-201.
Stehman, S. V., & Czaplewski, R. L. (1998). Design and analysis for thematic map accuracy assessment: fundamental principles. Remote Sensing of Environment, 64(3), 331-344.
Hi Bibi - Its good to take as much as possible points for high accuracy and the number of pixels in one point must be n(n+1), here n is number of classes.
Number of sampling have a positive relation with accuracy in case of stratified random sampling. But always here is a problem for field sampling and image classification is that sampling error. when we try to increase our sampling there is possible to sampling error. so good and accurate sampling also an important issue on accuracy of classification.
I determine sample but is a problem for small classes. there are lower than 20 point in small classes. i can increase it but morn's i= it is min distant will not be consider.
You can use an equation to calculate how many samples you need, it's very basic :
n = z * (P.q/e²)
n: number of samples
z: 2 * DS ( DS is the standard deviation)
P: the expected accuracy ( it's experimental it can never be 100 %, the ideal you should aim for is btw 85% - 95% )
q : (1-P)
e: the allowed error percentage of your accuracy assessment ( if z = 85% => e = 15%)
in that way you can calculate the average samples that are required for your accuracy assessment, however sometimes the number might be high and you don't have the necessary resources you should consider minimizing it.
still, now there is no clear-cut point and consensus regreding the sample size per area and per land use/cover. The objective of the analysis, resolution of the imagery, and others significantly determine the number of points. anyhow the more sample size is the more accurate.