This is impressive, Absolutely love this kind of projects!

Keep going :)  

Abdul

I believe that your problem is overtly generic in its formulation. To believe that a set of sensors will 'solve' most of the problems posed by traditional agriculture is a bit naïve. Real time data does not necessarily help a farmer or a company, as it bring a new set of problems, the most apparent being how to understand and pre-process and analyse the datasets. If I were you, I would break it into smaller problems that are easier to define and solve; for example, in a region with increasing drought problems, irrigation sensors will feature more heavily than others. Another issue is the reach, quality and duration of your sensors in collecting data. A capacitive soil sensor will not give you any indication about mineral content in the soil, it simply measures resistance. Another massive problem is sensor calibration; if you are measuring things like irradiance, reflectance, or even more complicated things like fluorescence, the calibration process is painful and labour intensive.

Let's talk for one second about your quadcopter, but focusing just on the payload. that is, I don't even want to get involved into the huge challenge of deploying a drone remotely, as this is the realm of serious engineering. So, let's talk about the camera, you mention two in fact, one nir camera and one RGB. I asume that both consist of the same sensors and lenses; I also asume that you do some postprocessing with the images such as mosaicing and image registration. For these two types of processes, you need much more than just images, you need accurate and relevant metadata, not only precise GPS coordinates, but also the camera angle, altitude, timestamp at millisecond level, and, OF COURSE, calibration data. The easiest way of getting calibration data is by deploying a lambertian pannel that you need to photograph prior, and after your mission, so you have a meaningful way of processing you images. Otherwise, all these pictures are useless. You can also collect that data the hard, very hard, way. It means you need to develop an up-down welling sensor, preferably with spectrometers, but it can also be built using photodiodes that measure individual bands. Such sensor would have to be mounted on a sort of gimbal on top of your drone. Otherwise, all the angles, turns, twists, etc. will destroy your data. Another option is to collect such data using a base station but, then again, that data will differ of what you camera is collecting and will need to be cleaned using complex maths. Let's talk about image sensors. You need to use expensive, monochrome sensors. Why expensive? Well, you want to be the proud owner of a product sheet with all the specificacions, such as QE. Not all sensors are sensible to light in the same way and if your camera is good capturing nir light, it might not be that good capturing blue light, so, you will have to acount for that in your postprocessing. Why monochrome? A bayer pattern filter will mess your data collection simply because is a layer printed on top of your sensor and it will change the amount of photons that reach in each color. You can compare the brightness of an image collected with a real monochrome sensor and one of the bands of your rgb camera, and you will see that there is a big difference. Of course, you can always compensate for that but, again, you one your data to be accurate as possible.

Then let's talk about ML and processing of images to gain indexes and substract from your images things like types of crops,n weeds, or even more complicated, nutrient deficiencies. Everyone relish the wholy NDVI as an everything-solver. It is not. Fro training models to be accurate, e.g. supervised classification, etc. you need good data. Data that you will need to collect by hand on the field. The data you need are spectral profiles of things you'll look for on you images later. So you need a spectral profile of a specific weed, spectral profiles of the crops you want to discern, etc. To collect spectral data you need a spectrometer calibrated for reflectance, and a tonne of field days. And hey! I am just talking about how things would look like. What about looking for plants with a specific problem that you want to find later? A disease, a nutrient deficiency, water stress? How do they differ of other things? Fever can be a sympton of tonnes of illnesses, so if you looking for fever, how do you know what is causing it?

If you speak to a farmer, or an industry specialist, they will be extremely sceptical of all-out solutions; sometimes they would even frown at the sight of 'problems' they don't experience, or simply are not high on their most do list. For example, by the time a nutrient deficiency is visible in your imagery, there might not be solution anymore, so the solution is to prevent these things to happen in first place. If you farmer is looking for nutrient deficiencies during the season, his crop is already gone.  

Another issue for field sensors is that they interfere with traditional agriculture machinery. Farmers need to spray and fertilize their fields several times in a season. if your sensors are bulky or tall, they will send you packing because they don't want to have more issues.

I recall a conversation with the CEO of a large UK based palm oil producing company with plantations in Indonesia. I told him we could tag each individual plant with RQ codes and the field workers will use their phones to log info about each plant: fruit weight, diseases, prunning, etc. He laughed and told me that his company was barely coping with all the data they already had, and more data will mean more confusion for his field staff. I left the meeting fuming, but he was deadly right.

Please do not let me discourage you. I hope the comments above are food for thought. I bet that you hhave solved most of these issues already. Best of luck with you project!

@Mayke Thank you so much for your insights. I am relatively new to this field and after you pointed out the loopholes in the solution I did lots of background research and discussed some issues with my professors.

I also took help from a startup founded by a senior of mine based in India. You were absolutely right about your concerns. I cannot thank you more. It is people like you who uplift the spirit of open-source. I will keep updating the logs as I move on with the project.

I was not familiar with the helium network, seems cool. Wouldn't lorawan and https://www.thethingsnetwork.org/ have a better range and be better suited for the big area of a farm?

I get your concern but please also take into account the cost of deploying LoRa nodes which is way beyond poor farmers can afford. LoRaWAN is obviously better for long distance applications but the setup eliminates the need for LoRa